fix(jit): widen kernel IDs and preserve full cache hashes

[jinzhen-lin]

Kernel registration can fail when distinct cubin paths share a CRC32 and the kernels have the same symbol name, as seen in Buildkite build 61. Compute a shared nonnegative 63-bit registration ID from CRC-64/ECMA-182 over the full cubin path, a null separator, and the kernel name. Apply this to both GEMM and process-input registration, retain collision checks, and initialize the lookup table safely under concurrent calls.

Return all 32 hexadecimal MD5 characters from hash_to_hex instead of truncating to 16. This changes cache keys, so affected JIT and native artifacts will be rebuilt.

Validation:

• python -m pytest tests/test_jit_infrastructure.py tests/kernels/process_input/test_scale.py -q — 107 passed on RTX PRO 6000 Blackwell GPUs.
• CRC64 standard vectors, concurrent initialization, field separation, and a known CRC32 path collision checked with a temporary standalone C++ program.
• Full MD5 known vectors, Ruff, formatting of changed C++ code, and git diff --check.