Styled underlines and underline colour, and the sub-parameters they need

[JohnCampionJr]

SGR 4:3 asks for a curly underline — how an LSP marks an error. It was arriving as parameter 43.

The cause is bigger than underlines

The parser transitioned to CsiIgnore on a colon and abandoned the whole sequence. So 38:2::255:0:0 — the ITU-T spelling of a foreground colour, which plenty of programs emit — produced no colour and no error. Params.AddSubParam existed and nothing called it; GetSubParams was a stub returning an empty list, comment included:

// Sub-parameters are stored contiguously
// This is a simplified version
return result;

So most of this PR is sub-parameter support, and styled underlines are what it unlocks.

Sub-parameters are stored flat with a start offset per parameter — almost no sequence has any, and the ones that do have a handful, so a list per parameter would allocate on every CSI to describe nothing. An empty slot is kept as a real zero: dropping it would shift 58:2::255:0:0 by one and turn red into black.

One trap worth flagging for review. A colon case written beside the digit branch in the state machine can never be reached, because 0x3A sits inside the 0x30..0x3F parameter-byte range that branch already claims. I wrote it there first and it silently did nothing — which may be how the original gap went unnoticed. It belongs in Param(), next to the semicolon.

What lands on top

• SGR 4:0–4:5 select the underline style; 21 is a double underline
• SGR 58 sets the underline's own colour, in both the colon and semicolon spellings; 59 restores it to the foreground
• IsUnderline() now reports the style rather than a separate flag, so the two can't disagree

The UnderlineStyle enum already existed and already matched Ghostty's — none/single/double/curly/dotted/dashed.

It costs nothing

AttributeData.Extended holds nine flags in a 32-bit field, so the style went in bits 9–11 and an interned colour id in 12–31. The struct is still twelve bytes and the cell still 24 bytes, and there's a test asserting that rather than trusting it.

Interning was chosen over two alternatives, both measured rather than argued:

• Growing the struct — a bigger cell costs most on fills; going 24 → 32 bytes cost scroll-heavy output 22% in earlier work.
• Interning the whole style, as Ghostty does — that needs reference counting, which forces every cell write to read the old id to release it. Measured at 240 ns per line against 165. An underline colour releases nothing, so it adds nothing to the write path; the lookup happens once per run per frame on the render side.

Full RGB is representable. What's bounded is how many distinct colours coexist — about a million, against the four an LSP uses.

Tests

18 new, covering every style, both 58 spellings, indexed and RGB, reset behaviour, interning identity, that the new bits don't disturb the other attributes, and that the struct didn't grow.

867 passing. Independent of #32, #34 and #35 — parser and attributes, no buffer. The renderer side (drawing curly, dotted, dashed) is Iciclecreek work and I'll open it separately.

[JohnCampionJr]

@tomlm I also tested this against the perf work and there was no affect so thats good. anything we do going forward should run before/after tests to make sure we don't accidentally chip away at the hard won gains of that PR