fix(profile): retry a failed read; require reversibility for unattended writes

Testing the round trip turned up a defect the unit tests could not: a
single transient failure of `read_current` silently converts a reversible
write into a one-way one. The write still succeeds, `previous` is None,
and nothing says so until someone tries to undo it. Seen for real --
`PowerGetActiveScheme` failed once while several compiles were running,
then read normally on the next attempt, which is also why the fix is a
retry rather than a redesign.

`read_current_with_retry` gives a flaky source three chances.
`apply_setting_reversible` refuses outright when the prior value still
cannot be read, and `tuning::serve` uses it for every unattended write: a
loop with nobody watching must not make a change it cannot take back.
`apply_setting` keeps the old behaviour for attended callers, who can be
told the write is one-way and decide for themselves.

Tests added while hunting this:

- read -> parse -> format round-trips for the Windows GUID, including the
  all-zero and all-ones edges. This is the seam a revert crosses and it is
  invisible if only one side is tested: `read_current` formats with Debug,
  `apply` parses with `parse_guid`, and nothing forces them to agree.
- malformed GUIDs are rejected rather than truncated into a valid-looking
  different one.
- 2000 concurrent reads across 8 threads, because `read_current` allocates
  and frees Win32 memory on every call and the tuning loop calls it every
  cycle; a mistake there is heap corruption showing up somewhere else.
- the retry itself, with a handler that fails once and one that always
  fails.
- an ignored end-to-end test that applies, verifies, reverts and verifies
  again against the real machine. Ignored because it writes; run with
  `cargo test --all-features --lib -- --ignored round_trip`.

Verified by hand beyond the suite: the round trip moved this machine
Balanced -> High performance -> Balanced, `powercfg /getactivescheme`
independently confirmed it was returned to where it started, and the audit
log carries `previous` on both entries, so a revert is reconstructible
after a restart rather than only in-process.

860 tests, 6 consecutive clean runs of the lib suite (4,404 executions, no
flake), clippy -D warnings clean, release build and docs checked.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>