Remove init from validate, add init to set_strategy, introduce init_strategy to interpolators

[kylecarow]

Resolves #20.

Lifecycle fix

Interpolator::validate took &mut self only because it re-ran strategy.init(&self.data)
as a side effect, duplicating what new already does and papering over a real gap:
set_strategy swapped the strategy but never initialized it, so a stateful custom
strategy swapped in at runtime was silently left uninitialized.

• validate is now a pure &self check: data shape/monotonicity and the extrapolate
  setting only. It no longer touches the strategy.
• set_strategy (all Box<dyn Strategy*D> and *Enum overloads, across
  Interp1D/Interp2D/Interp3D/InterpND) now calls the new strategy's init
  against the current data, mirroring new.
• Added init_strategy(&mut self) to each concrete interpolator type as a discoverable
  way to re-run that same init step directly. Useful after mutating the public
  data/strategy fields, or after deserializing an interpolator with a stateful
  custom strategy, since Deserialize doesn't go through new/set_strategy and
  won't call init on its own.
• Updated examples/custom_strategy.rs's comment describing when init runs.
• New tests: set_strategy surfaces a strategy's own init error (using Step's
  direction-count check), and validate()/init_strategy() diverge as expected when
  data is mutated directly in a way that breaks LinearUniform's uniform-grid
  requirement (validate() passes, init_strategy() catches it).
• README: expanded the "Validation Lifecycle" section to describe validate's
  narrowed scope and point at init_strategy for the strategy-specific case.

This is a breaking change to the Interpolator trait signature, consistent with #20's
proposal to land it in the next pre-1.0 minor bump rather than a patch.

Unrelated cleanup

While in these files, fixed everything cargo clippy --all-features --all-targets and
cargo doc --all-features --no-deps flagged:

• 6 broken rustdoc intra-doc links ([Interpolator]/[Extrapolate] in lib.rs and
  strategy/traits.rs) now use full paths so they resolve.
• 16 needless_borrow warnings in the interpolator/*/tests.rs serde round-trip tests.
• 5 needless_range_loop warnings in interpolator/n/strategies.rs's bitmask/butterfly
  reduction loops, rewritten with .iter().enumerate() / .zip() (same math, no
  behavior change).
• 2 missing_transmute_annotations warnings in examples/uom.rs, given explicit
  transmute::<ArrayView1<Ratio>, ArrayView1<f64>> turbofish instead of relying on
  inference.

Verified locally: cargo fmt --check, cargo hack test --feature-powerset, all
examples (default and --all-features), cargo doc --all-features --no-deps,
cargo clippy --all-features --all-targets, and cargo bench --no-run are all clean.

CI

Added clippy and doc steps to .github/workflows/test.yaml (cargo clippy --all-features --all-targets -- -D warnings and cargo doc --all-features --no-deps
with RUSTDOCFLAGS=-D warnings), so the class of warnings fixed above is caught going
forward instead of silently accumulating. Both pass clean on this branch.

Changelog

Introduced a project changelog.