fix(ci3): run dedicated benches on bare metal (m6a.metal) to cut variance (#24298)

## Problem

The dedicated bench box (`m6a.32xlarge`) still produced bimodal
benchmark numbers — e.g. `avm_bulk` wire commitments flipping between
~2.1s and ~3.2s — despite CPU isolation, HT-off, and (separately tried)
NUMA memory binding.

## Root cause (confirmed by a live controlled experiment)

On a fresh `m6a.32xlarge` I ran `avm_bulk` (a) alone and (b) while 3
more `avm_bulk` saturated the other NUMA nodes:

| Phase | isolated | contended (all 64 cores) | Δ |
|---|---|---|---|
| Wire commitments (MSM) | ~2,070 ms | ~2,098 ms | +1.4% |
| Sumcheck (compute) | ~2,770 ms | ~2,873 ms | +4% |

Saturating the whole box barely moved the numbers — so it's **not**
code, **not** contention from our own concurrent benches, and **not**
NUMA locality (memory binding made no difference). Across every observed
run the **compute phases are constant** while **only the
memory-bandwidth-bound MSM phase varies box-to-box**. That's the
signature of **per-physical-instance memory bandwidth differing**:
`m6a.32xlarge` is ~2/3 of a host and shares it with other tenants, so a
memory-heavy co-tenant steals bandwidth, and placement is random per
launch → the bimodal flip.

## Fix

Run the dedicated bench box on **`m6a.metal`** — sole-tenant bare metal,
no co-tenant to contend for memory bandwidth. Unchanged: spot-first with
on-demand fallback (same sole-tenant hardware either way; mid-run
reclaim handled by `bootstrap_ec2`'s `NO_SPOT` retry). Bonus: 96
physical cores vs the 32xl's 64, so more bench parallelism.

Supersedes #24276 (the NUMA-binding attempt, now closed — the experiment
showed binding wasn't the lever).

Author: charlielye

ci.sh

@@ -180,17 +180,21 @@ case "$cmd" in
     ;;
   bench)
     # Launched by the build instance on uploadable runs to produce stable benchmark
-    # numbers on a dedicated instance of a FIXED type. AWS_INSTANCE pins the exact type
-    # (bypasses spot pool diversification) — that's what keeps numbers comparable. Spot
-    # vs on-demand is the same hardware, so we try spot first and fall back to on-demand
-    # (the default fleet behaviour); a mid-run spot reclaim is handled by bootstrap_ec2's
-    # internal on-demand retry. CI_DASHBOARD and PARENT_LOG_ID are inherited from the
-    # launching run so it nests as a sibling job.
+    # numbers on a dedicated, FIXED-type instance. AWS_INSTANCE pins the exact type
+    # (bypasses spot pool diversification) — that's what keeps numbers comparable. We use
+    # bare metal (m6a.metal): a sub-host instance (e.g. m6a.32xlarge) shares its physical
+    # host with other tenants, and a memory-heavy neighbour steals memory bandwidth, which
+    # showed up as bimodal variance in the bandwidth-bound proving phases (MSM commitments)
+    # while compute phases stayed flat. Metal is sole-tenant, so there's no neighbour.
+    # Spot vs on-demand is the same hardware, so we try spot first and fall back to
+    # on-demand (the default fleet behaviour); a mid-run spot reclaim is handled by
+    # bootstrap_ec2's internal on-demand retry. CI_DASHBOARD and PARENT_LOG_ID are
+    # inherited from the launching run so it nests as a sibling job.
     # Timestamp the instance-request output. pipefail (in a subshell, since ci.sh doesn't
     # set it globally) keeps bootstrap_ec2's exit code through add_timestamps — the
     # launching build instance waits on this fatally.
     ( set -o pipefail
-      AWS_INSTANCE=m6a.32xlarge JOB_ID=x-bench INSTANCE_POSTFIX=x-bench \
+      AWS_INSTANCE=m6a.metal JOB_ID=x-bench INSTANCE_POSTFIX=x-bench \
         bootstrap_ec2 "./bootstrap.sh ci-bench" 2>&1 | add_timestamps )
     ;;
   socket-fix)