#2093 sm_75: the prompt path's top-k with one full read of the scores (identical ids)

Beschreibung

**Builds on #2072** (it sits next to that PR's split top-k in `qsa_select.cu`); the first commit here is #2072's, the second is this one.

Past ~115K cells the prompt path's top-k is DRAM traffic on Turing: `block_topk_wide_kernel` reads a query's scores six times (four radix passes, a count, the emit), 400 KB a query at 400K, 256 queries a call.

`block_topk_sample_kernel` reads them once:
1. a bound LB: the KS-th largest key among every 32nd block (KS = 2 × the selection's blocks / 32 + 16);
2. one read of all blocks, keeping those with key ≥ LB in shared memory;
3. when those hold at least `width` cells, the width-th largest cell key (block_topk_kernel's threshold) is ≥ LB, so every block at or above the threshold is a candidate: the threshold, `above` and the tie budget computed over the candidates with block_topk_kernel's rule are the reference's, and the candidates sorted by block (bitonic) emit the reference's cells in its order;
4. otherwise (too few cells above LB, or more than 4,096 candidates) the query is flagged and `block_topk_wide_kernel<true>` runs it - the all-equal-scores case of the parity test takes that path.

Used for prompt calls above 28,672 active blocks (below, the wide kernel is as fast); `STRATA_TOPK_SAMPLE=0` keeps the wide kernel.

## Setup

RTX 2080 Ti 22 GB (sm_75), i7-6700, CUDA 13.0, driver 595, SM 1650 MHz, 240 W; `v0.1.43` + #2072, `-DSTRATA_BUILD_TESTS=ON`. Model Q2_0, `--max-context 524288`, `--kv int8 --kv-resident 32768`, `STRATA_SELECT_SIMT=2` (#2071).

## `qsa_topk_parity CTX 256 20 524288 1` (a prompt call), dispatched time

| context | wide | sample |
|---|---:|---:|
| 98K | 0.397 ms | 0.396 (wide: below the bound) |
| 131K | 0.513 | 0.452 |
| 262K | 0.952 | 0.589 |
| 400K | 1.476 | 0.706 |
| 520K | 1.840 | 0.787 |

`qsa_topk_parity --selftest`: PASS (continuous, 64-level and equal scores; identical ids). `qsa_topk_active_parity` (`STRATA_TOPK_ACTIVE_ANY=1`): 20 cases, identical.

## Engine

First logits after a 196K prompt: bitwise identical to the run without it (`STRATA_DUMP_FIRST_LOGITS`). The gain sits at depth: with this, #2071 and #2091 and #2092, appending 16K tokens to a 390K conversation (`--serve`) went 15.9 → 12.9 s here.

🤖 Generated with [Claude Code](https://claude.com/claude-code)