ci: spec22-gems msys harness presses probe images in the leg's format

[ronaldtse]

Why

The second arm64 dogfood (tebako-runtime-ruby run 35690023046) got past staging and the bridge tree with #112, then failed at the probe presses:

== spec22-gems-msys step: press setup image (GemRunner shim)
Error: mkimage failed: mkimage --format dwarfs is not available on windows-arm64
       (the dwarfs-t writer has no arm64 windows closure yet); use --format limnifs

All three mkimage calls hardcoded --format dwarfs. On windows/arm64 the runtime is limnifs-only by design (the link unit has no native dwarfs-t under clangarm64), so the driver under test never mounts dwarfs bytes and the CLI refuses the press by name.

What

PRESS_FORMAT derives from the same arch detection #112 added (MSYS_TOOLCHAIN): dwarfs on x64 (the x64 env image's format), limnifs on arm64. The setup image and both payload probes (materialized + negative oracle) share the one derivation, so the proof legs always mount the format the runtime under test ships.

Proof

• bash -n clean; the only dwarfs mentions left are the new comment and the x64 arm of the derivation.
• Exact error string quoted from the failing run's log; the fix follows the CLI's own named-error guidance.