internal: add tuple struct support in pin_data and pin_init! (V2)

[mqqz]

Replaces #113.

This adds tuple struct support to #[pin_data], init!, and pin_init!.

The projected form of a tuple struct is also a tuple struct, so projected fields are accessed with native tuple syntax (.0, .1, ...), which keeps the generated API closer to
ordinary Rust and avoids exposing synthetic field names (also makes it easier fields are cfg'd out).

• tuple struct projections from #[pin_data]
• tuple struct initialization with indexed brace syntax:
  • pin_init!(Foo { 0 <- ..., 1: ... })
• tuple struct initialization with constructor syntax:
  • pin_init!(Foo(...)) (but no <- e.g. Foo(a, <- b, c) is rejected)
• support for tuple structs with generics, const generics, pinned drop, and fallible partial initialization rollback
• support for tuple structs when fields or constructor arguments are removed by #[cfg] (Error out unless #[cfg] is on the final tuple field)

Notes

• tuple projections use native tuple field access rather than synthetic names e.g. ._0 like last time.
• cfg handling does not try to evaluate user #[cfg] conditions in the proc macro
• instead, the macro generates the necessary cfg-dependent layouts and lets rustc select the active branch
• #[cfg] is supported only on the final tuple field / constructor argument, when it does not change numbering.

Tests

Added coverage for:

• basic tuple struct projection
• brace and constructor tuple initialization syntax
• duplicate / missing / invalid tuple field UI failures
• generic and const-generic tuple structs
• #[pin] interaction with !Unpin field types
• pinned drop
• fallible partial-init rollback
• cfg-stripped tuple fields (last or otherwise)
• cfg-stripped tuple constructor arguments
• feature-dependent cfg field layouts

Comparison with the old PR

This replaces the earlier attempt in #113.

Compared with that PR, this version is intentionally narrower and easier to review:

• less code motion/churn overall (my code always got motion 😎)
• fewer refactors mixed into the feature work
• tuple projections are tuple structs instead of relying on synthetic internal/public field names
• cfg handling was redesigned to correctly handle cfg-stripped tuple fields and constructor arguments simplified to accept only on last field/arg. or error out
• the test set was trimmed to the non-redundant cases, while keeping the important regressions and restoring the fallible partial-init rollback coverage

Closes: #85