Normalize scalar tensors in CuTe JIT cache keys

[xyzhang626]

CuTeDSL compile keys are compared and hashed on the Python side. If a scalar tensor gets into the key, two keys with the same scalar value but produced by different layers compare by object identity and miss both the in-memory and persistent JIT caches.

This can happen in FA4 varlen backward when max_seqlen_q/k are scalar tensors and the backward compile key includes derived single-block predicates such as seqlen_q_rounded // m_block_size == 1. Each layer can produce a fresh scalar tensor, causing repeated backward JIT compilation for the same kernel shape.

This PR normalizes scalar tensors in CuTe JIT cache keys before in-memory lookup/storage and before computing the persistent cache hash.

Tests added:

• in-memory cache lookup with distinct scalar tensor objects but identical values
• persistent cache hash stability for scalar tensor keys

Local validation:

• manual CPU scalar cache normalization check passed
• manual CUDA scalar cache normalization check passed on GPU7

Could not run pytest locally because pytest is not installed in the available environment.

[xyzhang626]

Just realized this some overlaps with #2507. #2507 fixes the specific bwd max_seqlen source, while this PR normalizes scalar tensors at the JIT cache boundary so future scalar tensor compile-key entries do not miss the in-memory or persistent cache.