Fast GPU ASR v0.2.0

Maintenance release improving decoding, concurrent inference stability, and test coverage.

• Improve Parakeet TDT beam search with full-vocabulary and duration-aware hypothesis merging before pruning.
• Fix cuFFT plan-readiness ordering in Parakeet and Zipformer feature plugins for concurrent execution.
• Support feature-plugin workspaces larger than 2 GiB for larger batches and longer audio, while retaining overflow checks.
• Reduce peak GPU memory during encoder context resizing by releasing the previous allocation before allocating its replacement.
• Improve word-timestamp handling for SentencePiece unknown, control, and byte-fallback tokens.
• Refine CUDA device and stream handling, with expanded runtime and plugin tests.
• Reduce ONNX export logging noise while preserving warnings and errors.

The high-level ASR interface remains unchanged. Direct users of low-level encoder/decoder classes now pass a CuPy device object; postprocessing accepts audio durations instead of waveforms.

Full changelog: v0.1.0...v0.2.0