v0.6.0

Follow-up to 0.5.0: makes a documented setting real, and removes one that was
pretending to work.

A minor rather than a patch release because the greplm-core library API changes.
For anyone depending on that crate:

• the io-uring cargo feature is gone;
• config::Backend::IoUring is renamed IoUringRemoved (it now resolves to the
  portable backend);
• io_backend::UringBackend is gone.

Nothing changes for CLI or MCP users, and existing config.toml files keep
working — backend = "io-uring" is still accepted, with a warning.

Changed

• backend in config.toml now actually selects the ingest backend. It was
  parsed and validated, then ignored — selection came from build features alone,
  so setting rayon or io-uring changed nothing. greplm status also reports
  the backend genuinely in use rather than the one a build feature implied.

Removed

• The io-uring cargo feature and its backend. It was a stub that deferred to
  a buffered read while reporting itself as io_uring, so a build with the feature
  on claimed a backend it did not have. Measuring the ingest path across the worker
  pool on 5,161 kernel files with a warm page cache: tree-sitter parsing 92.2%,
  reads 5.6%, hashing and trigram extraction 2.1%. A perfect batched-read backend
  could therefore remove at most ~5.6% of per-file CPU, and less in wall clock
  because reads already overlap parsing. The IoBackend trait stays as the seam
  and docs/ROADMAP.md records the numbers so the decision can be revisited for
  cold-cache indexing of very large trees.

  A config.toml written before this change may still say backend = "io-uring";
  it keeps working, resolving to the portable backend with a warning, rather than
  failing every command over a value that never did anything.

Full Changelog: v0.5.0...v0.6.0