Add DRY sampler stage (SS_DRY)

[sashko-zakharchuk]

Adds the DRY (Don't Repeat Yourself) repetition penalty as SS_DRY, the dry_multiplier half
of #270 (the logit_bias half is #274). TabbyAPI currently warns "Ignoring sampler params not
supported by the exllamav3 backend: dry_multiplier"; I'll open the companion TabbyAPI PR to pass
the parameters through once this lands.

SS_DRY penalizes a token that would extend a sequence already present in the context by
multiplier * base ** (match_length - allowed_length), so verbatim loops get exponentially
expensive while short incidental repeats are untouched. It is modeled on SS_RepP: it runs
among the first steps on raw logits, participates in the fused-sampler fast path as a leading
step, no-ops when dry_multiplier == 0 or dry_base < 1, and is wired into ComboSampler as
dry_*. The scan is pure torch on whatever device holds past_ids, with no host-device sync,
so there is no extension change.

The matching semantics and the float32 exponent-overflow clamp follow llama.cpp's
llama_sampler_dry_apply (the koboldcpp implementation by pi6am, scheme by p-e-w), including
never penalizing breaker tokens and the base = 1.75 / allowed_length = 2 defaults. The
parameter surface follows exllamav2: dry_multiplier, dry_base, dry_allowed_length,
dry_range (0 = whole context, where llama.cpp's equivalent defaults to 64), and
dry_sequence_breakers as token IDs (TabbyAPI's breaker strings resolve to IDs with the
exported dry_sequence_breaker_tokens). Left as None, the breaker set derives from the
tokenizer at sampling time with exllamav2's charset rule
(get_dry_default_sequence_breaker_tokens).

Divergences, also stated in the docstrings:

• exllamav2's own DRY is a different algorithm (an incremental trie that weights the penalty by
  occurrence count and caps matches at dry_max_ngram = 20). This port follows llama.cpp
  instead, since that is what circulating DRY presets are tuned against. If you would rather
  match the exl2 behavior, I can rework it.
• Sequence breakers are single tokens. llama.cpp additionally builds multi-token restart
  sequences out of partially overlapping tokens; that machinery is not ported.
• Match lengths are capped at 2048 tokens. Where dry_allowed_length plus llama.cpp's exponent
  clamp exceeds 2048 (dry_base below ~1.044 at the default dry_allowed_length), a verbatim
  repeat longer than the cap is penalized as a 2048-token repeat. The cap is what keeps the
  scan bounded for bases near 1.

Cost, measured on an RTX 5050 with a 128k vocabulary: 0.5 ms per sampled token at 1k-8k token
windows with the default dry_base = 1.75, 6.6 ms at a 128k-token window. A small base widens
the scan up to the 2048 cap (dry_base = 1.1 gives 2.6 ms at 8k, 37 ms at 128k). Transients
are chunked, ~200 MiB worst case. The docstring says to cap dry_range for very long contexts.

Tests: ten deterministic cases (window bounds, breaker capping and exclusion, batch rows, the
float32 exponent-clamp boundary), a randomized differential against a brute-force oracle on
both CPU and CUDA past_ids (200 cases each, covering chunk-boundary crossings and
out-of-vocabulary IDs), default-breaker derivation through a stub tokenizer, and fused-collapse
assertions. Outputs were also cross-checked against the DRY port in vllm#50584: 1200/1200
randomized cases match. Sampler suite: 151 passed vs 137 on master, and clean with
EXL3_FUSED_SAMPLER=0.