Cache NPUKernel objects

[mawad-amd]

This PR implements function-level caching for JIT-compiled kernels using a global cache that stores compiled kernels based on function signatures, argument types and contents (for source code). This avoids redundant compilation by reusing cached kernels when the same function is called with identical argument types and shapes.

Example:

(ironenv) ➜  00_memcpy git:(muhaawad/cache-objects) ✗ rm -rf ~/.iron && python memcpy.py

Latency: 0.002165 seconds (2165.37 µs)
Effective Bandwidth: 61.98 GB/s

PASS!

(ironenv) ➜  00_memcpy git:(muhaawad/cache-objects) ✗ python memcpy.py 
Latency: 0.002128 seconds (2127.71 µs)
Effective Bandwidth: 63.08 GB/s

PASS!