Add Python frontend for Neura hardware custom ops

[YanzhouTang]

Summary

• Introduces python/neura_ops.py as the Python frontend interface for Neura hardware primitives.
• Programmers can explicitly call custom ops (e.g., torch.ops.neura.gather()) to annotate hardware-specific memory access patterns that standard PyTorch ops cannot express.
• Custom ops behave identically to standard PyTorch operations at runtime, but are preserved as-is through torch_mlir tracing so that the Neura compiler backend can recognize and lower them to hardware primitives.

Example: neura::gather in hash encoding

• hash_encode kernel uses indirect addressing (level_embeddings[indices]) for hash-table lookups. After torch_mlir tracing, this becomes aten.index.Tensor, which loses the gather semantics needed by CGRA.
• With the custom op, programmers write torch.ops.neura.gather(table, indices) instead. torch_mlir preserves it as torch.operator "neura.gather", giving the compiler a clear signal.
• Refactored trilinear_interpolation to batch all 8 corner indices into a single gather per level (from 8 calls down to 1).

Changes

1. python/neura_ops.py (new): Registers neura::gather with CPU and Meta implementations, compatible with PyTorch 2.1+.
2. nerf_kernels.py (new): Imports neura_ops, replaces fancy indexing with torch.ops.neura.gather(), two-phase trilinear interpolation.
3. compile_hash_encode.py (new): Compilation script that exports clean ~930-line Torch Dialect MLIR with neura.gather preserved.

Verification

• neura.gather appears exactly 2 times in output MLIR (one per level).
• aten.index.Tensor fully eliminated.
• PyTorch-side correctness verified (semantically equivalent to the original implementation).

[tancheng]

neura.gather appears exactly 2 times in output MLIR (one per level).

I don't see MLIR output?

PyTorch-side correctness verified (semantically equivalent to the original implementation).

The correctness is based on "semantic", instead of execution, right?

[YanzhouTang]

neura.gather appears exactly 2 times in output MLIR (one per level).

I don't see MLIR output?

MLIR output is generated by running compile_hash_encode.py

PyTorch-side correctness verified (semantically equivalent to the original implementation).

The correctness is based on "semantic", instead of execution, right?

Current correctness verification is semantic-level. Maybe we could register neura.gather in the Neura interpreter, which will enable execution-level correctness verification at the MLIR layer?

[tancheng]

neura.gather appears exactly 2 times in output MLIR (one per level).

I don't see MLIR output?

MLIR output is generated by running compile_hash_encode.py

PyTorch-side correctness verified (semantically equivalent to the original implementation).

The correctness is based on "semantic", instead of execution, right?

Current correctness verification is semantic-level. Maybe we could register neura.gather in the Neura interpreter, which will enable execution-level correctness verification at the MLIR layer?

Then can we have a lit test to run the .py and check its output?

neura.gather in the Neura interpreter

We can do this later (another PR once your design/infra is almost done).