[KernelGen][Nvidia] Add in-place sign_ operator

[ShawnsYing]

Summary

Adds the in-place sign_ operator to complement the existing sign / sign.out
operators that were introduced in #5030. sign_ writes the element-wise sign
back into the input tensor.

aten::sign_ has only the default overload (sign_(Tensor(a!) self) -> Tensor(a!)),
mirroring the abs / abs_ pairing. The implementation reuses the already-merged
_sign_impl kernel with out == input, so no new kernel logic is introduced.

Changes

• src/flag_gems/ops/sign.py: add sign_(x) → _sign_impl(x, x)
• src/flag_gems/ops/__init__.py: export sign_
• src/flag_gems/__init__.py: register ("sign_", sign_) in _FULL_CONFIG
• conf/operators.yaml: add id: sign_ block
• tests/test_sign.py: test_sign_ (all real dtypes + shapes) and a
  non-contiguous case; asserts the mutated input matches torch
• benchmark/test_sign.py: test_sign_inplace

Testing

191 pytest cases pass against this worktree (isolated launcher verifying
flag_gems resolves to the worktree). Dispatch confirmed via GEMS SIGN_
debug log; verified same storage pointer (true in-place) and NaN → 0.
sign_ supports all real dtypes including int / uint8 / bool; complex raises
NotImplementedError.

[ShawnsYing]

KernelGen SimpleOpt result

Ran KernelGen SimpleOpt against the sign trace on NVIDIA H20 (3 rounds, all PASSED):

Round	geo_mean	Notes
1	1.032x	Baseline: element-wise Triton kernel, tl.where for sign, BLOCK_SIZE=1024 (best)
2	regressed (-22% on large 1D)	Branch-free arithmetic via bool→float conversion
3	regressed	BLOCK_SIZE=4096 exceeded the 1024-thread HW limit, destroying occupancy

Best: 1.032x geo_mean speedup vs the PyTorch reference.

sign is a trivial memory-bandwidth-bound pointwise op, so the straightforward Triton
implementation is already near-optimal against PyTorch's highly-tuned reference — ~parity
is the expected result for this class of unary op. The in-place sign_ added here reuses
the same merged _sign_impl kernel (out == input), so it inherits this performance.

[bin913]

/test:sign_:h20

[github-actions]

⚠️ Cannot proceed with operation

• reviewDecision: skip_reviews
• commitStatus: PENDING

Reviews are not required for this operation but CI checks must be passing in order to continue

[bin913]

/test:sign_:h20

[github-actions]

❌ On-demand test failed

Operator: /test|sign_
Runner: h20
Backend: nvidia-cuda133

The test failed to complete. Check the workflow run for details.

[bin913]

/test|sign_:h20