chore: stop xtask from bloating target/ and slowing the gate

[mohamedmansour]

The local quality gate (cargo xtask check) was taking 18+ minutes on long-lived working copies, even though the actual compile + test work was only ~2 minutes. The root cause was the workspace target/ directory growing to 160+ GB / 1.1M+ files from accumulated incremental compilation state and per-feature fingerprint variants. Cargo then spent most of every subsequent invocation just stat'ing fingerprint files.

Two changes:

• Force CARGO_INCREMENTAL=0 on every cargo invocation that xtask spawns. xtask never runs an inner edit-compile loop — every cargo call it makes (gate steps, examples build, bench, E2E, packaging) is one-shot, so incremental compilation just produces garbage that cargo then has to scan on the next run. CI was already immune to this via Swatinem/rust-cache@v2 (which sets the same env). Real inner-loop development with cargo build / cargo test outside xtask is unaffected.

• Switch the gate's bench-validate step to --profile=dev. The step exists only to run criterion's --test smoke path, but the default bench profile inherits the release profile's lto = true, codegen-units = 1, so it was recompiling the entire microsoft-webui graph with full LTO just to assert criterion's entry point loads. Dev profile reuses unit-test artifacts and drops that step from ~44s to ~2s. Real benchmark measurements via cargo xtask bench still use the unchanged bench profile, so reported numbers remain accurate.

Measured on a working copy where target/ had grown organically:

before (bloated target/, 160 GB):  18m23s
after  (bloated target/, 160 GB):  N/A — change forces clean rebuild
after  (clean target/):             42s   (first run)
after  (clean target/):             35s   (subsequent runs, no growth)

bench-validate step alone:

before:  ~44s  (bench profile, LTO + codegen-units=1)
after:   ~2-3s (dev profile, reuses test artifacts)

[mcritzjam]

Dev prod ftw