fix(flashlfq): bound each peak window to one charge state

[trishorts]

@Alexander-Sol — this targets PeakEnvelopeOutput, so it lands inside #1122 rather than alongside it. It's the one blocking item from my review there; the other four findings are design calls I've left to you.

The problem

ChromatographicPeak.IsotopicEnvelopes holds one envelope per scan per charge state — that is what NumChargeStatesObserved counts. PeakWindow.Create took a single Min/Max of IndexedPeak.M across all of them, so a peptide traced at more than one charge got one window spanning the gap between the charge states.

Measured on PeakEnvelopeOutput before this change, with EGFQVADGPLYR (1350.66 Da) at z=2 (676.3 m/z) and z=3 (451.2 m/z), over a synthetic MS1 with one peak every 0.5 m/z:

NumChargeStatesObserved = 2
Window    = 450.700 - 676.800 m/z   (width 226.1 Da)
PeakCount = 1356 over 3 scans       (452 peaks/scan)
TSV rows for this ONE peak = 1356

[same peptide, single charge] width 1.0 Da, PeakCount = 6, rows = 6

226x the rows, and essentially none of that 226 Da is co-elution around the precursor — it's the rest of the spectrum. On real Orbitrap MS1 data (10-50k peaks over 350-1500 m/z) the output is effectively unbounded. I suspect the 200 Da window you benchmarked in the PR description is exactly this rather than a pathological case.

The change

• PeakWindow.Create now takes the chargeState whose envelopes bound the window, and only those envelopes set MinMz/MaxMz, the scan span, and the peakfinding flags.
• PeakWindow.CreateForEachChargeState yields one window per traced charge state, ascending.
• PeakWindow.ChargeState is exposed, and the Peak Charge column now carries the window's charge state rather than Apex?.ChargeState ?? 0.
• WritePeakWindows writes every charge state of a peak under that peak's single peakId, distinguished by Peak Charge. The correspondence with QuantifiedPeaks that the comment there promises is unchanged.
• ApexRetentionTime is now the apex of the window's own charge state. The peak's overall apex can belong to a different charge state whose scan lies outside the window entirely, which would have put an out-of-range RT in the RT Apex column.
• Peak RT Start / Peak RT Apex / Peak RT End renamed to Window RT *, since they describe the window's scan span and that span is now per charge state.

Tests

Your 13 tests still pass, updated for the new signature. Two added:

• TestPeakWindowIsBoundedPerChargeState — the regression above. Builds the two-charge peak over a spectrum with a peak every 0.5 m/z, asserts two windows in ascending charge order, each exactly 2 * mzExpansion wide and centred on its own m/z, and that the total peak count stays under 50 where the old bounding box swept in ~1356.
• TestPeakWindowApexIsPerChargeState — z=2 apexes in the first scan and z=3 in the last; asserts each window reports its own apex rather than the peak's.

TestWritePeakWindows additionally asserts every row's Peak Charge is a charge the peak was actually traced at, and that every traced charge state contributed rows.

Also added BuildMultiChargePeak, so a test can build the peak FlashLFQ actually produces for a multiply-charged peptide; BuildPeak now delegates to it and is unchanged for callers.

Verification

• Solution builds clean, x64|Debug, 0 errors.
• All 15 tests in TestPeakWindow pass.
• Patch coverage over the whole of feat(flashlfq): export the raw MS1 peaks around each quantified peak #1122 plus this change, Coverlet/Release against upstream/master: 98.99% (196/198 changed executable lines), target 90. No file below the bar.

Take it, adapt it, or close it — whichever is least friction for you.