ci(fpga): place gf6/gf8/gf16-mul-clean and gf24-sub with HeAP, not SA

[gHashTag]

Second batch from the #862 placer audit, after #863. Four workflows move from --placer sa to --placer heap. Nothing else changes: same image tag, router1, --freq 50, seeds 1-8, first clean seed wins.

Measured in run 37215621928 (exp-sa-audit-ab2), which used the same sources, xdc and Yosys flags as each workflow, and seeds 1-16 for both placers:

workflow	SA clean of 16	SA, this loop today	HeAP clean of 16	HeAP s/seed	Fmax SA → HeAP
ax7203-gf16-mul-clean	4 (6, 11, 12, 15)	98 s, seed 6	16	8-10	27.9-37.2 → 60.6-66.5 MHz
ax7203-gf8-mul-clean	4 (4, 10, 11, 12)	80 s, seed 4	16	5-7	42.3-51.4 → 102.2-116.0 MHz
ax7203-gf6-mul-clean	4 (8, 10, 11, 15)	194 s, seed 8	16	5-6	42.4-59.1 → 146.6-159.2 MHz
ax7203-gf24-sub	3 (6, 8, 13)	215 s, seed 6	16	6-9	15.6-16.2 → 26.3-28.1 MHz

"This loop today" is the sum of the per-seed times for seeds 1 to the first clean one, i.e. what the existing step spends before it breaks. Under HeAP it stops at seed 1.

Why SA failed, over the four designs: 49 of 64 seeds failed. 41 of those hit post-placement validity check failed for Bel 'SLICE_.../A5FF' (openXC7/nextpnr-xilinx#162; 40 say "(no cell)", one names a cell), 6 left $PACKER_VCC_NET unrouted, and 2 hit the router1.cc:331 assert valid_wires_for_net.count(w). HeAP: 0 of 64 failed.

Also fixed: the three MUL workflows named their step "placer heap + per-seed timeout" while passing --placer sa. The name is now true.

Left alone, for the owner: ax7203-gf16-conformance.yml (gf16_codec). Same picture: SA 2 of 16 clean (seeds 8 and 9), HeAP 16 of 16 at 2-4 s, Fmax 83.7-95.4 → 258.3-309.8 MHz. Its loop only goes green because seed 8, the last seed it tries, happens to place. It prints Fmax into the log, so switching it changes a number someone may be quoting; that is a separate decision.

Not measured

• The workflows pull the unpinned regymm/openxc7 tag; the A/B pinned sha256:35c91073. This PR's own CI runs are the check that the tag still behaves the same.
• The bitstreams from these workflows have not been on the board under either placer.
• Nothing here touches the router. router2 is a separate question (fpga: measure nextpnr router2 / --threads now that place-and-route dominates #850).

Refs #862

🤖 Generated with Claude Code

[github-actions]

X Brain Health Check

Score: 100.0/100
Status: 🟢 HEALTHY
Threshold: 80/100

X Brain is above merge threshold

[github-actions]

🧪 Brain Stress Test Results

Score: not measured (no stress test prints a Score: line)
Status: 🟢 PASS
Threshold: 270/300

X Stress test passed