Implement missing Sampler combinations

[stikves]

The CompositeSampler running on the Accelerate / CPU currently supports

• TopK
• TopP
• Temperature

(Along with optimized path for Greedy, Temperature=0)

However the MPSGraph based counterpart used in pipelined engine only supports

• TopK
• Temperature

(Again a similar fast path for Greedy)

We should be able to implement TopP, however that requires a scatter operation.

Similarly we are missing potential upgrades for

• Repetition penalty
• MinP (cheaper alternative to TopP)

But those are not higher priority

Though the high effort / high return item is

• Adding GPU based constrained sampling

As this forces guided generation paths to use sequential engine.

[stikves]

Reference:
https://blog.mlc.ai/2026/05/04/xgrammar-2-fast-customizable-structured-generation