Skip MTP probes while drafts go unused

[pandysp]

Loading --mtp makes every plain (non-speculative) eval run the MTP drafter, and when the next call is also a plain eval the draft is invalidated unused. Sampled decoding does this on every token, so any temp>0 span pays a full drafter pass per token with no way to use the result. Measured on an M4 Max (q2-q4-imatrix base, Q4K/Q8 MTP head), paired and interleaved: generating at temp 0.7 with --mtp loaded runs 6.5% slower [6.3, 6.7] than without it. The tax is easy to miss because any baseline with --mtp loaded pays it too, including DS4_MTP_SPEC_DISABLE runs.

The fix is one session counter: when a pending draft is invalidated unused, skip the next 15 probes; a consumed draft resets it. Greedy decoding never arms the skip, since its drafts are always consumed, so the speculative path is untouched: greedy output is byte-identical with and without the patch on copy and prose prompts, and --mtp-verify-depth passes.

With the patch, sampled generation with --mtp loaded is back to within 0.4% [0.1, 0.8] of not loading it. Under a live pi coding-agent task against ds4-server (mixed greedy tool-call spans and sampled spans, 250k ctx), token-weighted decode improves +2.7% to +7.6% over three paired rounds (mean +5.4%), with the task succeeding in every run.

Not addressed, measured for scope: low-acceptance greedy content (free-flowing prose at temp 0) still pays the probe for drafts that are consumed but rejected; that is the -5.4% draft-2 row in the #369 table, a different effect from this tax. A rate-based gate prototype recovers about half of it; glad to send that separately if you want it, kept out of here so this stays one mechanism.