Add fused reduction support

[sakupan102]

Closes #4355.

• Introduced the CompiledReduce primitive, which fuses unary element-wise operations
  with a following all-reduce operation.
• Added JIT generation and execution of fused reduction kernels with unary prefixes
  for Metal and CUDA.

Metal results

Function and input	Before (us)	After (us)	Change	Speed-up
max(abs(x)), 32 x 32	5.610	5.225	-6.9%	1.07x
max(abs(x)), 128 x 128	14.780	9.544	-35.4%	1.55x
max(abs(x)), 1024 x 1024	111.615	35.655	-68.1%	3.13x
sum(exp(abs(x))), 1024 x 1024	111.318	35.064	-68.5%	3.17x

TODO

• Support column and row reductions. The current implementation supports only
  all reductions.
• Support fusing into ArgReduce.

---

• ☑️ I understand it is strictly prohibited to use AI to write PR description
• AI usage disclosure: I used Codex for writing the code.

[sakupan102]

@nastya236 Could you take a look at this?

[sakupan102]

@nastya236
Since this is quite a large change, would it be better to split this PR into multiple smaller ones (e.g., separate PRs for Metal, CPU, and CUDA support)