fix: reduce retained geometry memory in large reviews

[matthew-hre]

Hello! Been a big fan of Hunk recently, I use it basically every day at work -- however, it's been a bit of a memory-hog on my machine, so I thought I'd start doing some digging and hopefully contribute some performance patches! First contribution on this project, let me know if I missed anything.

This PR introduces a fix that reduces memory usage in large reviews by making diff section geometry hold on to less data.

Previously, DiffSectionGeometry was holding on to the full planned row stream for every measured file section -- most review interactions really only need lightweight geometry data: row bounds, hunk bounds, body height, and lookup maps.

Since we already window the rendered rows, keeping full planned rows for the entire changeset means we're still retaining a bunch of row objects, even when most files weren't mounted.

This changes geometry to keep the bounds/navigation data eagerly, but rebuild planned rows lazily, only when copy selection asks for row content.

I also added a focused bench:geometry-memory benchmark so this path is easier to measure next time.

Results

The biggest improvement showed up on the huge synthetic review benchmark:

Metric	Main	This PR	Difference
first-frame heap	228.5MB	175.1MB	-53.4MB / -23%
after-navigation heap	324.4MB	248.9MB	-75.5MB / -23%
after-navigation RSS	708.8MB	644.3MB	-64.4MB / -9%

The resize memory benchmark also improved, which makes sense:

Metric	Main	This PR	Difference
final heap	149.7MB	120.0MB	-29.6MB / -20%
heap growth	110.2MB	88.0MB	-22.1MB / -20%
peak RSS	481.2MB	453.1MB	-28.2MB / -6%
RSS growth	216.6MB	186.8MB	-29.8MB / -14%

Little less impressive on the memory.ts benchmark, but still positive:

Metric	Main median	This PR median	Difference
first-frame heap	35.6MB	31.3MB	-4.3MB / -12%
after-navigation RSS	481.2MB	465.4MB	-15.8MB / -3%

I also ran the new focused geometry benchmark on this branch:

bun run benchmarks/geometry-memory.ts
METRIC geometry_heap_growth_bytes=7469892
METRIC geometry_rss_growth_bytes=5898240
METRIC materialize_planned_rows_ms=6.17
METRIC materialized_planned_rows=10260
METRIC materialized_planned_rows_heap_growth_bytes=0

This should be fairly cut-and-dry. Rendering still builds planned rows for mounted sections, and copy selection still gets planned rows when it needs them. The only difference really is that cached geometry no longer keeps those rows alive for every file by default.

[greptile-apps]

PR author is not in the allowed authors list.