LowestFee::bound panics on assert!(ideal_fee >= 0.0) due to f32 rounding

## Summary

`LowestFee::bound` can panic via `assert!(ideal_fee >= 0.0)` on valid inputs, reachable through the public `run_bnb`/`bnb_solutions` API.

https://github.com/bitcoindevkit/coin-select/blob/cd8cec4/src/metrics/lowest_fee.rs#L276

## Cause

In the not-yet-funded arm of `bound()`, the "perfect resized input" relaxation computes:

```rust
let ideal_fee = scale.0 * to_resize.value as f32 + cs.selected_value() as f32
- target.value() as f32;
assert!(ideal_fee >= 0.0);
```

With sat values in the millions (well within f32's 24-bit integer mantissa limit of ~16.7M, but large enough that intermediate rounding bites), `scale = remaining / value` rounds such that `scale * value + selected - target` lands a tiny amount **below** zero. At `target.fee.rate == 0` this is easiest to hit, since `scale` is derived purely from the value gap.

## Reproduction

Found by differential fuzzing (10/1500 random scenarios with realistic UTXO sizes). One concrete panic case:

- 15 candidates with values up to ~1.5M sats, e.g. `{4167(w360), 13447(w7597), 1480489(w2663), 17281(w368), 26718(w755), 237889(w4570), 273139(w997), 145266(w2124), ...}`
- `Target { fee: { rate: 0.0 spwu, absolute: 0, replace: None }, outputs: { value_sum: 52692, weight_sum: 208, n_outputs: 1 } }`
- `LowestFee { long_term_feerate: 0.59 spwu, dust_relay_feerate: 1.4 spwu, drain_weights: { output_weight: 342, spend_weight: 815, n_outputs: 2 } }`
- `cs.run_bnb(target, metric, 100_000)` → `panicked at 'assertion failed: ideal_fee >= 0.0'`

## Suggested fix

`0` is always an admissible lower bound for a fee, so clamping preserves correctness:

```rust
let ideal_fee = (scale.0 * to_resize.value as f32 + cs.selected_value() as f32
- target.value() as f32)
.max(0.0);
```

(Alternatively compute the relaxation in f64 and keep a `debug_assert!` with an epsilon.)

Found while stress-testing #53; the bug predates that PR and reproduces on current master (cd8cec4).

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Start in src/metrics/lowest_fee.rs at LowestFee::bound, especially the not-yet-funded arm and its ideal_fee calculation. Reproduce the panic through the public run_bnb or bnb_solutions API using the provided scenario, then verify that valid inputs no longer panic and that the lower bound remains correct.

Written by the indexing model from the issue text.