Add GPU-based constrained sampling for pipelined engine

[stikves]

Enables grammar-constrained generation (JSON schema) on the pipelined engine by applying the xgrammar bitmask inside the MPSGraph sampler on GPU, eliminating the forced fallback to the slower sequential engine.

Changes:

• Add bitmask expansion graph to MPSGraphArgmaxSampler and MPSGraphCompositeSampler (bitwise AND + notEqual + cast, zero measurable latency overhead)
• Add ConstrainedGenerationSession.fillBitmask(into:) for zero-copy write into shared MTLBuffer
• Add CoreAIPipelinedEngine.generateConstrained() with semi-pipelined loop (inference overlaps bitmask computation)
• Add PipelinedConstrainedDecodingStrategy for the GPU path
• Route pipelined engine to GPU strategy in CoreAILanguageModel

Performance:

Bitmask expansion adds 0.65ms per token (measured on 32K vocab). On a Qwen3-VL-2B model, constrained generation on pipelined vs sequential produces identical outputs with comparable throughput

Tests:

• Added new unit tests, and verified existing ones
• Ran end-to-end inference
• Greedy outputs are identical between pipelined and sequential