fix: accept * N anywhere in the track header, not only the first slot

[spacedevin]

The bug

* N was recognized only in the slot immediately after gen <id>. Anywhere else it fell through to the key/value loop and became a macro param literally named *:

track Pad id pad gen gameBoyDmg layer 0 * 16

→ loopBars: null, genParams: { layer: 0, "*": 16 }

The pattern length silently became 1 bar instead of 16. No error, no warning — just a song playing at 1/16th of its intended length.

How it was found

Moving tish-gba onto this parser (it had its own hand-written one that scanned header tokens in any order). I diffed the baked ROM data for all 77 shipped .deck files, old parser vs new — 75 were byte-identical and 2 were not. Both were this.

It's not rare in that corpus:

`* N` immediately after gen (canonical):   174 track lines
`* N` after other params (misparsed):       24 track lines, 4 files

This survived because emit always writes * N first — so round-tripping through a host never reproduces it. It only bites hand-written and generated files, which is most of tish-gba's corpus.

Fix

* N and the trailing key value pairs are now accepted in any order after gen <id>. And a * that names no valid length (* 0, * zz, a bare trailing *) is an error rather than a quietly dropped token — a pattern length that silently becomes 1 is the worst kind of wrong.

canonical       loopBars=16  genParams={"layer":0}
star last       loopBars=16  genParams={"layer":0}
star middle     loopBars=4   genParams={"layer":0,"cutoff":900}
inf last        loopBars=null
star bad        errors=["track * N: expected number or inf/infinite"]
star trailing   errors=["track * N: expected number or inf/infinite"]

Verification

Corpus case 002-track-header gains both non-canonical orderings, so the expected-parse file now pins this. Suite green, coverage back at 100% lines/functions/statements (54/54).

Once this releases and deckfile republishes, tish-gba's 77-song bake becomes byte-identical to its old hand-written parser — which is the acceptance test for retiring that parser entirely.