[jiapingW]

Adds an independent eager softmax-attention backend under layers/attention/backends/. It consumes prepared Q/K/V tensors and an explicit mask, scale, dropout probability, and training flag. It owns no projections, architecture settings, learned parameters, or causal/window policy.

The backend supports MHA/grouped KV heads, unequal query/key lengths, and distinct query/value widths. Boolean masks use True for visibility; floating masks are additive, with negative infinity blocking a key. Fully blocked rows return zero with zero gradients instead of NaNs. Finite negative biases retain their additive meaning. Invalid tensor/mask layouts fail before attention computation.

Validation on H20-wjp:

• 7 backend tests passed. CPU FP32/BF16 and CUDA FP32/FP16/BF16 outputs and Q/K/V/bias gradients match HF Qwen3 eager attention exactly in the tested ordinary-row cases, including training dropout, broadcast masks, noncontiguous inputs, and grouped heads.
• CPU/GPU fully blocked rows agree with the PyTorch SDPA math reference (rtol 1e-5, atol 1e-6); blocked outputs and their gradients are exactly zero. Additional tests cover block/context isolation, empty query/key lengths, and early validation.
• 231 configuration tests and 36 existing component/HF-loading/backbone tests passed. The parent PR's 27 rotary tests and both PRs' pre-commit checks also passed.

Uses the existing Transformers 5.12.1 environment; the project's pinned 5.19.0 environment was not exercised. This supplies the numerical backend for subsequent attention components; it does not yet implement/register GQA projections, build algorithm masks, or migrate model entry points. No serving/performance claim is made. Stacked on #1121; one commit.

Historical PR for original Draft design 41. Superseded by current Draft design 06/21, #1124. Its code is included in that PR; this historical PR remains closed.