[CI] Drop the duplicate gfx950 test runner and revert the nightly README badges

[coderfeli]

Two small CI cleanups, both on top of main.

Drop the duplicate gfx950 runner

linux-flydsl-mi355-1 and linux-flydsl-mi35x-1 are the same target:

Runner	runner-config.yml	parse_bench.py arch
linux-flydsl-mi355-1	MI355, 1 GPU	gfx950
linux-flydsl-mi35x-1	MI350, 1 GPU	gfx950

Every push therefore ran the identical RUN_TESTS_FULL=1 scripts/run_tests.sh
twice on identical hardware — see run 33710068450,
where both were still going long after mi325 and navi had finished. The same
duplication existed for the -8 pair in the multi-gpu matrix.

This drops the mi35x entries from test, test-skip, multi-gpu
(flydsl.yaml) and the wheel test matrix (test-whl.yaml).

Why keep mi355 and not mi35x: the main branch ruleset requires
test (linux-flydsl-mi325-1) and test (linux-flydsl-mi355-1).
test (linux-flydsl-mi35x-1) is not required, so keeping mi355 means no
branch-protection change is needed and there is no window where PRs stall
waiting on a check that is no longer reported.

Why not race the two pools: these are ARC scale sets, not labelled
self-hosted runners, so runs-on: must name one exactly. There is no shared
label that would let GitHub hand the job to whichever pool is free, and both
matrix jobs dispatch simultaneously with no way to observe each other.
Statically dropping one entry is the only available option.

Revert the nightly README badges

Reverts only the README.md hunk of #1091, restoring the badge row to
CI / Benchmark / Dashboard / Docs. The three nightly integration workflows
themselves are untouched and still run on schedule; this only removes the
links. The prepare-mlir manylinux changes from that commit are left fully
in place.

Verification

• Both workflow files parse as YAML.
• The test and test-skip matrices remain identical (they must be, or
  docs-only PRs hang on the required checks).
• Both required contexts are still emitted.
• README.md is byte-identical to its pre-[CI] accelerate prepare-mlir: run the CPU builder in the manylinux image #1091 state.
• The diff touches no manylinux / prepare_runner lines.

🤖 Generated with Claude Code