subset capi: expose configurable input sets, flags, and preprocessing

[behdad]

Extend the subset C API beyond Skia's initial glyph workflow with hr_subset_input_set, hr_subset_input_keep_everything, and hr_subset_preprocess, plus HarfBuzz aliases. Expose configurable names/languages, layout features/scripts, and dropped-table sets, along with the implemented Skera flags for legacy names, overlap bits, unknown-table passthrough, glyph names, Unicode ranges, layout closure, and bidi closure.

These APIs are part of the optional subset feature of harfrust_capi; enable it with cargo build -p harfrust_capi --features subset and link the single harfrust_c library. Headers, examples, and tests live under harfrust_capi.

Preserve inverted input tag sets when building Skera plans. Skera uses that representation to select existing script records efficiently; materializing all possible tags would make it enumerate billions of values. Preprocessing returns an immutable standalone face owning copied tables, without promising an acceleration cache.

Published Skera fixes its no-subset-table set internally, so modifying that set currently returns failure rather than ignoring the request. A small upstream Plan API extension is being prepared for customization and glyph mappings. Axis instancing, CFF/CFF2/VARC, and unimplemented dependency flags remain unavailable.

Validation: the original 48 differential cases and 15 additional HarfBuzz comparisons pass, including input-set defaults, keep-everything, name/layout/drop controls, all added flags, unknown tables, overlap bits, and preprocessing ownership. Rust 1.85 checking, rustfmt, and Clippy pass. The CI C smoke test now also exercises preprocessing, inverted selections, and subsequent customization.

Stacked on #511; extends #507.