Currently, the GHA jobs take around 45' to complete, what really slows down the development cycle. In this PR I introduce a number of improvements to the main integration test workflow.

First, the cpp-funcs and py-funcs steps only need to be re-run if there are any changes in the clients/cpp and clients/python submodules respectively. Unfortunately, the upload-artifact action only works within a workflow run, and not accross workflow runs, so we can't use previous run's artifacts easily. Instead, we use a cache for WASM code. In summary, the cpp-funcs and py-funcs only run if there are changes or the cache has expired. (As a sidenote, GHA's caches have a versioning that depends on the compression algorithm used, which is different inside and outside a container, so we must make sure they are the same by installing zstd in the containers).

Secondly, a lot of time is spent building the codegen binary without sanitisers, and then re-building the tests again with sanitisers enabled. We make a couple of observations:

• First, code generation needs to happen only if (i) cpp or python have changed, or (ii) cpp and python have not changed, but we have a machine code cache miss (i.e. new CPU).
• We only need to build the codegen_func binary if we need to re run the codegen, and not always.