Partially revert #14410. That PR altered `Kind::is_number` and `Kind::is_any_keyword`, replacing `matches!` on `Kind` variants and method calls with explicit matching on the enum discriminant range e.g. `matches!(self as u8, x if x >= Decimal as u8 && x <= HexBigInt as u8)`.

The rationale for the change was improved performance. I was suspicious of the AI's claims that this was a perf improvement, as I assumed that compiler would boil down the code to the same anyway.

I asked Claude to generate the ASM before / after and he concluded it makes no difference either way.

> The compiler generates identical code - a simple sub + cmp + setb range check in both cases.
>
> ```asm
> // is_number
> add dil, 107
> cmp dil, 11
> setb al
> ```
>
> LLVM recognises that:
>
> 1. Contiguous matches! variants can be reduced to a range check.
> 2. The union of the 4 keyword sub-methods covers a contiguous range, so it merges them into a single range check.
>
> The commit's claim about "20% fewer instructions" and "better pipeline utilization" is nonsense. The manual `as u8` casts just make the code harder to read for zero benefit.

New Claude pretty damning about old Claude!

So go back to explicit matches, as it's easier to understand.