[fast-launcher] Concrete-tensor fast path skips _hashable_dims (Python-only)

[yushangdi]

Stacked PRs:

• Speed up Helion kernel launches by avoiding repeated Python work #2565
• ->[fast-launcher] Concrete-tensor fast path skips _hashable_dims (Python-only) #2611
• [fast-launcher] Codegen rewrite to wire the pool into generated wrappers (Python-only) #2605
• [fast-launcher] Opt-in output-tensor pool (Chunk D, Python-only) #2604

---

[fast-launcher] Concrete-tensor fast path skips _hashable_dims (Python-only)

Splits the per-tensor specialization-key extractor in two:

• _concrete_tensor_key (new) handles torch.Tensor and
  torch.nn.Parameter. Both always have concrete-int sizes/strides,
  so the cache-key components are just obj.size() and
  obj.stride() directly — torch.Size and the stride tuple are
  already tuple subclasses whose hash matches a plain tuple of the same
  ints, so the dispatch produces an identical hashable key without
  rebuilding it.
• _tensor_key (existing, unchanged) stays as the
  FakeTensor-dispatch entry. FakeTensors have SymInt sizes during
  tracing, and _hashable_dims is the normalization that maps each
  SymInt to (id(shape_env), expr) so two SymInts from different
  shape envs don't accidentally collide.

The split happens at the _specialization_extractors table, so the
right extractor is selected by a single dict lookup with no per-call
type check on the hot path.

Hash equality between the two extractors is verified by a test (see
test_fast_path_key_hash_matches_wrapped) so existing on-disk
LooseAutotuneCacheKey entries built before this change keep
matching.

Benchmarks (H100, vector_add at N=4096, cycling through 8 different
x/y tensors so the cache hits but never on the same args twice)

_base_specialization_key in isolation (the dominant component of
Kernel.bind() per-call cost):

before: 2.76 us
after: 2.14 us
saving: -0.62 us (-22%)

End-to-end per-call timing at this stack position (Chunk D pool +
codegen rewrite applied, Phase 2 fast launcher not yet applied):

before (parent commit, debc08a): cpu_only=~16.0 us wall+sync=~16.0 us
after (this commit): cpu_only= 14.87 us wall+sync= 14.87 us
saving: ~-1.1 us/call

Tests

test/test_tensor_key_fast_path.py covers:

• Dispatch table routes concrete tensors to the fast path and
  FakeTensor to the old path.
• The fast-path key hashes and compares equal to the old wrapped key
  (cache-compat invariant).
• The fast-path key contains a torch.Size (not a wrapped plain
  tuple) when the kernel uses static_shapes=True.
• bind() still returns the same BoundKernel for different
  tensor objects with matching dtype/shape/stride, and distinguishes
  different dtypes and shapes.