Support suppress_tokens / begin_suppress_tokens (HF generation parity)

[Copilot]

onnxruntime-genai had no way to suppress specific token IDs during generation, so output diverged from transformers.generate() for models that depend on suppress_tokens / begin_suppress_tokens (e.g. Whisper, Gemma). These options were neither parsed from config nor applied as logits processors.

This adds static per-token suppression mirroring HF's SuppressTokensLogitsProcessor and SuppressTokensAtBeginLogitsProcessor:

• Config (config.h/config.cpp): new suppress_tokens / begin_suppress_tokens int-array fields on Config::Search, parsed via a new Search_Element::OnArray.
• Logits processor (search.h, search.cpp, cuda/search_cuda.cpp): new Search::ApplySuppressTokens virtual, implemented for CPU and CUDA, setting targeted logits to float lowest — following the existing ApplyMinLength / ApplyRepetitionPenalty pattern.
• Application (generators.cpp, engine/request.cpp): suppress_tokens applied every step; begin_suppress_tokens only when current_length == prompt_length. Wired into both the standalone Generator loop and the batching Engine request loop.
• API surface: additive C API OgaGeneratorParamsSetSearchTokensArray + C++ wrapper; Python set_search_options(...) now accepts list[int].
• Model builder (builders/base.py, builders/whisper.py): auto-populates both from generation_config.json / model config via a shared Model.add_suppress_tokens_to_search_config helper, so existing HF models work without manual config edits.
• Tests: C++ greedy-selection tests and Python API tests covering every-step and begin-only suppression.

# Suppressed at every step; greedy falls through to the next unsuppressed token
params.set_search_options(suppress_tokens=[1, 3], begin_suppress_tokens=[50257])

bad_words_ids (multi-token sequence banning), listed as optional in the issue, is intentionally out of scope here.

Note on concurrency

The begin-step is tracked via a per-object length member, mutated non-atomically — consistent with the existing single-threaded-per-Generator/Request design (cf. computed_logits_, is_prefill_). No synchronization added.