refactor!: Replace Selector with a single into_selection pass

[evanlinjin]

Description

Reworks the coin-selection API around a single InputCandidates::into_selection pass and cleans up its error surface.

The public Selector struct is removed as nothing used it. Its methods were only ever called in sequence by into_selection, so it was public surface with no consumer. That logic now lives inside into_selection, and algorithms take a bdk_coin_select::CoinSelector plus a pure-data SelectionContext instead of a &mut Selector.

This also:

• unifies the two redundant long-term feerates (change policy + bnb metric) into one SelectionParams::longterm_feerate;
• splits the catch-all SelectorError so each error type only carries variants it can actually produce;
• fixes a reachability bug that produced a false CannotMeetTarget under RBF.

Notes to the reviewers

• Behavior change (please scrutinize): the change policy is now always waste-aware (min_value_and_waste). Previously a None long-term feerate selected a plain dust-only policy. Consequence: raw-script change must now supply a real satisfaction_weight (descriptor change is unaffected — it's derived). Rationale: min_value_and_waste with longterm == target does not collapse to dust-only; it still charges the change output's lifetime cost, which is the correct thing to weigh.
• Bug fix detail: the reachability pre-check called select_all_effective(raw target_feerate) but scored with excess(target.fee.rate) (the RBF-resolved, higher rate). This could result in an under-selection. Fixed by selecting at target.fee.rate.
• Error split: to_cs_change_policy previously returned a SelectorError that included a LockTypeMismatch it could never produce. It now returns ChangePolicyError (only Miniscript / InsufficientAssets), and LockTypeMismatch is a top-level IntoSelectionError variant.

Changelog notice

Added:
- Add `SelectionContext`, passed to selection algorithms alongside the `CoinSelector`.
- Add `ChangePolicyError` (`Miniscript`, `InsufficientAssets`) as the return type of `SelectionParams::to_cs_change_policy`.
- Add `IntoSelectionError` variants `ChangePolicy`, `LockTypeMismatch`, `CannotMeetTarget { missing }`, and `AlgorithmFellShort`.

Changed:
- Rename `SelectorParams` to `SelectionParams`.
- `InputCandidates::into_selection` no longer takes a `&mut Selector` closure; algorithms are now `FnOnce(&mut bdk_coin_select::CoinSelector, SelectionContext) -> Result<(), E>`.
- Replace `SelectionParams::change_longterm_feerate` with `SelectionParams::longterm_feerate`; `None` now means "use the target feerate", and the change policy is always waste-aware. The single value feeds both the change policy and the lowest-fee-bnb metric.
- `selection_algorithm_lowest_fee_bnb` no longer takes a `longterm_feerate` argument (it reads it from `SelectionContext`).
- Fix a false `CannotMeetTarget` under RBF: the reachability check now selects effective inputs at the resolved target feerate.

Removed:
- Remove the `Selector` struct (its lifecycle is folded into `InputCandidates::into_selection`).
- Remove `SelectorError` (replaced by `ChangePolicyError` and dedicated `IntoSelectionError` variants).

Before submitting

• I followed the contribution guidelines
• This PR breaks the existing API

[noahjoeris]

Is this pre-check right for RBF cases as well?

select_all_effective(target.fee.rate) only keeps EV > 0. under RBF, could we still need a rate-negative utxo to hit the absolute fee floor?

[evanlinjin]

Good point, there should be an absolute value check as well.

[ValuedMammal]

The public Selector struct is removed as nothing used it. Its methods were only ever called in sequence by into_selection, so it was public surface with no consumer.

I didn't understand the stated rationale. Public items are for consumers of the library, and not required to be used internally.

[evanlinjin]

@ValuedMammal fair - that framing is incorrect and I'll fix it. Internal call sites aren't evidence about downstream consumers.

The actual rationale: the Selector was a wrapper that didn't provide much value. Algorithms now receive the downstream CoinSelector directly with a pure data SelectionContext. This is less code for us to maintain, and is actually more surface than what CoinSelector provided.