Fix positioned top-k sampling for speculative generation

[tjansn]

Summary

• apply top_k inside _PositionedTargetSampler, before temperature scaling
• keep seeded position-keyed sampling enabled when top_k is non-default
• add regressions for positioned-sampler selection and top-k filtering

Problem

Seeded speculative generation already uses a position-keyed target sampler so proposal draws cannot perturb target draws. The selection gate excluded every request with non-default top_k, however, and the positioned sampler did not apply top-k itself. That sent seeded top-k generation back through the stateful sampler path, where serial and speculative grouping can consume target randomness differently.

The server sampling path applies top-k before temperature/top-p sampling. This patch ports that same ordering to the positioned target sampler and removes the now-unnecessary top_k == 0 selection restriction.

Validation

• targeted selection/filtering assertions pass locally
• Python compilation passes
• a local Qwen MTP matrix covering two quantization widths, two seeds, opposite serial/speculative arm orders, and five sampling regimes passed 40/40 reports after the fix, including temperature=1, top_p=1, top_k=40
• seeded completions repeated across opposite arm orders; proposal draws did not perturb target draws

I could not run the full upstream pytest suite in this environment because pytest is not installed in the pinned runtime.