refactor(pytorch): replace backend operator builders with typed build specs

[grimoire]

Motivation

PyTorch operators were constructed through OpType and one-method builder
classes. This separated the operator identity, construction arguments, and
returned implementation type across multiple abstractions, making dispatch
harder to understand and extend.

Note

Since we have change the route of build_op, we better not merge this PR before other PRs that contain new ops.

Changes

• Introduce typed BuildSpec[ImplT] objects as the operator construction
  contract.
• Replace get_layer_impl_builder() with OpsBackend.build_op(spec).
• Migrate all PyTorch operator families, including:
  • attention and linear operators
  • quantized linear operators
  • fused MoE and activations
  • normalization, embedding, and RoPE
  • sampling and cache-copy operators
  • NSA, gated-delta, and DeepSeek-V4-specific operators
• Remove OpType and legacy one-method builder classes.
• Preserve Default, CUDA, and DLINFER provider selection, fallback behavior,
  and lazy imports.
• Preserve implementation ownership, weight loading, state-dict names, and
  public NN APIs.
• Reserve an internal enable_deterministic construction seam for future
  deterministic kernels. It is currently a no-op and is not exposed through
  the public configuration.
• Normalize related backend and build-spec naming.

This is a construction-time dispatch refactor. It does not intentionally
change operator forward behavior or kernel selection.