build: add a fast test profile to cut edit-test iteration from minutes to seconds

[inureyes]

Summary

[profile.release] in Cargo.toml is tuned for shipping binaries (fat lto = true, codegen-units = 1, opt-level = 3, strip = true). All local and agent test iteration currently runs through this profile because target/debug is usually cold (the MLX C++ tree rebuild makes the debug profile expensive to warm). Add a dedicated fast profile for test/dev iteration.

Motivation (measured)

Measured facts from the 2026-07-17 implementation chain:

• The main crate is a single ~390k-line crate, so any source edit recompiles the whole crate on one core (codegen-units = 1 blocks parallel codegen), followed by a fat-LTO link across ~439 locked crates. Measured at 4 to 6 minutes per incremental rebuild.
• A typical issue cycle (developer edit-test iterations plus review-stage fixes) pays 5 to 8 such rebuilds, i.e. 25 to 45 minutes of pure compile time per small issue.
• All-targets test runs multiply the LTO link per test binary.

Implementation plan

1. Add [profile.test-fast] to Cargo.toml:
   • inherits = "release"
   • lto = "thin" (or false if thin-LTO still dominates link time; measure both)
   • codegen-units = 16
   • strip = false
   • incremental = true
   • keep opt-level = 3 (MLX-heavy tests need optimized numerics for tolerable runtimes; drop to 2 only if measurements show parity)
2. Add a Makefile target test-fast (and optionally check-fast) wrapping cargo test --profile test-fast --features cuda, with a narrow-filter example in the help text.
3. Document in the contributor/build docs which profile to use for iteration vs shipping.
4. Parity check: run a representative narrow test set (e.g. models::minimax_m3, server::chat_request, mlxcel-core sampling and cache::ring) under both profiles and confirm identical pass/fail results; record the measured rebuild-time comparison (cold and incremental) in the PR body.

Acceptance criteria

• [profile.test-fast] exists in Cargo.toml and builds with --features cuda
• Makefile test-fast target exists and runs tests under the new profile
• Contributor/build docs updated to state which profile to use for iteration vs shipping
• Parity spot-check recorded: representative narrow test set produces identical pass/fail results under both profiles
• Measured incremental rebuild speedup reported in the PR (expect roughly 3 to 5x)

Note: repository CI currently runs no test job, so this is purely a local/agent iteration improvement with no CI changes required.