fix accepted-prefix consistency for streaming, history and stopping

[Gxj230958]

Lookahead can return one token sequence while streaming another: accepted block tokens are omitted, history repeats the last accepted token, and EOS/custom stopping inside a block is checked too late. This change determines the accepted prefix before KV-cache commitment and uses that prefix consistently for streaming, history, masks and return values in greedy and sampling paths.

Public augment_all() also required compatibility fixes to reach these paths: copy explicitly defined methods while preserving descriptors, retain Transformers constructors, make repeated augmentation safe, and preserve disabled sampling fallback. Default SDPA configuration now avoids an unsupported argument and transplanted zero-argument super(); legacy tuple-cache Lookahead falls back to manual attention, with no SDPA speed claim.

Validation on Torch2.7.1+cu128/Transformers4.36.2: 19 CPU tests and two GPU backend cases pass (each backend checks three random seeds, AR and Lookahead against the pre-injection HF reference). Tests exercise real tiny Llama forward passes, multi-token acceptance, multiple EOS IDs, internal stopping, streaming, history and cache lengths. The unmodified public augmentation baseline fails construction; a separately documented baseline with only compatibility fixes fails all11 decoding regressions. The new SDPA test also failed before its fix.

History behavior was reported in issue#62; disabled sampling fallback was reported in issue#54; the SDPA lookahead-kwarg TypeError was reported in issue#59 and is resolved by the SDPA configuration fix here. This does not modify the acceptance-count algorithm addressed by PR#68. FlashAttention2, multiple GPUs, newer Transformers and return-dict generation are not covered; the upstream return-dict restriction remains. FlashAttention2's ordinary HF Cache path still has a tuple-only assumption and is not claimed compatible.

Implementation and tests were prepared with AI coding assistance.