What does this PR do?
Partially restores #47007, which was fully reverted in #47083 due to
#incident-50242: only the absolute-path validation on XDG_CACHE_HOME is
reinstated.
In particular, the GOCACHE wiring that caused the incident is left out.

Motivation
A relative XDG_CACHE_HOME is silently accepted by the existing
directory-existence check and would be used as-is, leading to
cache misses or wrong paths depending on the working directory.

### What does this PR do?
Very partially restores #47007, which was fully reverted in #47083 due to #incident-50242:
- only the absolute-path validation on XDG_CACHE_HOME is reinstated,
- in particular, the GOCACHE wiring that caused the incident is left out.

### Motivation
A relative XDG_CACHE_HOME is silently accepted by the existing directory-existence check and would be used as-is, leading to cache misses or wrong paths depending on the working directory (see #47007)..

Co-authored-by: regis.desgroppes <regis.desgroppes@datadoghq.com>

### What does this PR do?
Reapply #47007, reverted in #47083 following #incident-50242, but with
the following adjustments:
1. scope repository caches to individual jobs (`$CI_JOB_NAME`) rather
   than runner types (`$CI_RUNNER_DESCRIPTION`),
2. add `.cache/ms-go*` to the cached paths (neutral until `msgo` 1.26+).

### Motivation
#incident-50242 traced to FIPS and non-FIPS omnibus jobs racing to write
incompatible Go build artifacts to the same cache key: `msgo` tags
indeed compile objects with `X:systemcrypto` while upstream Go does not.

It turns out to be a **known issue**: microsoft/go#1994.

Until the corresponding fix is available (microsoft/go#1998), that is
when bumping the Go SDK version to 1.26+, the main adjustment consists
in temporarily scoping repository caches to jobs.
Per-job cache keys naturally segregate the two without requiring an
explicit FIPS discriminator (we couldn't rely on `$FLAVOR` because it's
not expanding consistently across jobs).

### Additional Notes
When switching to Go SDK 1.26+, we'll revert `$CI_RUNNER_DESCRIPTION`
because `msgo` 1.26+ will eliminate the conflict at the toolchain level
thanks to microsoft/go#1998.

### What does this PR do?
Reapply #47007, reverted in #47083 following #incident-50242, but with
the following adjustments:
1. scope repository caches to individual jobs (`$CI_JOB_NAME`) rather
   than runner types (`$CI_RUNNER_DESCRIPTION`),
2. add `.cache/ms-go*` to the cached paths (neutral until `msgo` 1.26+).

### Motivation
#incident-50242 traced to FIPS and non-FIPS omnibus jobs racing to write
incompatible Go build artifacts to the same cache key: `msgo` tags
indeed compile objects with `X:systemcrypto` while upstream Go does not.

It turns out to be a **known issue**: microsoft/go#1994.

Until the corresponding fix is available (microsoft/go#1998), that is
when bumping the Go SDK version to 1.26+, the main adjustment consists
in temporarily scoping repository caches to jobs.
Per-job cache keys naturally segregate the two without requiring an
explicit FIPS discriminator (we couldn't rely on `$FLAVOR` because it's
not expanding consistently across jobs).

### Additional Notes
When switching to Go SDK 1.26+, we'll revert `$CI_RUNNER_DESCRIPTION`
because `msgo` 1.26+ will eliminate the conflict at the toolchain level
thanks to microsoft/go#1998.

### What does this PR do?
Reapply #47007, reverted in #47083 following #incident-50242, but with
the following adjustments:
1. scope repository caches to individual jobs (`$CI_JOB_NAME`) rather
   than runner types (`$CI_RUNNER_DESCRIPTION`),
2. add `.cache/ms-go*` to the cached paths (neutral until `msgo` 1.26+).

### Motivation
#incident-50242 traced to FIPS and non-FIPS omnibus jobs racing to write
incompatible Go build artifacts to the same cache key: `msgo` tags
indeed compile objects with `X:systemcrypto` while upstream Go does not.

It turns out to be a **known issue**: microsoft/go#1994.

Until the corresponding fix is available (microsoft/go#1998), that is
when bumping the Go SDK version to 1.26+, the main adjustment consists
in temporarily scoping repository caches to jobs.
Per-job cache keys naturally segregate the two without requiring an
explicit FIPS discriminator (we couldn't rely on `$FLAVOR` because it's
not expanding consistently across jobs).

### Additional Notes
When switching to Go SDK 1.26+, we'll revert `$CI_RUNNER_DESCRIPTION`
because `msgo` 1.26+ will eliminate the conflict at the toolchain level
thanks to microsoft/go#1998.

### What does this PR do?
Reapply #47007, reverted in #47083 following #incident-50242, but with the following adjustments:
1. scope repository caches to individual jobs (`$CI_JOB_NAME`) rather than runner types (`$CI_RUNNER_DESCRIPTION`),
2. add `.cache/ms-go*` to the cached paths (neutral until `msgo` 1.26+).

### Motivation
#incident-50242 traced to FIPS and non-FIPS omnibus jobs racing to write incompatible Go build artifacts to the same cache key: `msgo` tags indeed compile objects with `X:systemcrypto` while upstream Go does not.

It turns out to be a **known issue**: microsoft/go#1994.

Until the corresponding fix (microsoft/go#1998) is available, that is when bumping the Go SDK version to 1.26+, the main adjustment consists in temporarily scoping repository caches to jobs.
Per-job cache keys naturally segregate the two without requiring an explicit FIPS discriminator (we couldn't rely on `$FLAVOR` because it's not expanding consistently across jobs).

### Additional Notes
When switching to Go SDK 1.26+, we'll revert `$CI_RUNNER_DESCRIPTION` because `msgo` 1.26+ will eliminate the conflict at the toolchain level thanks to microsoft/go#1998.

Co-authored-by: regis.desgroppes <regis.desgroppes@datadoghq.com>