Speed up reload and resume

[dnouri]

Summary

Reloading and resuming long sessions should feel much snappier now.

This PR fixes the two slow paths that dominated profiling:

• large get_messages RPC responses are accumulated linearly instead of repeatedly copying the growing partial JSON line
• the resume picker no longer parses every session file twice, and metadata extraction avoids parsing full message payloads when it only needs counts/name/first message

It keeps one small render-path improvement: full history replay temporarily raises the GC threshold so GC does not interrupt the visible rebuild. The render benchmark is roughly flat; the big wins come from RPC framing and session metadata scanning.

It also adds a deterministic synthetic reload/resume benchmark suite with GUI/xvfb as the primary lane and batch as the CI-friendly smoke lane.

Benchmark results

Synthetic fixtures, GUI via xvfb, 3 repetitions per scenario, diagnostic timing advice disabled:

scenario	operation	before	after	speedup
large RPC response	resume	13.509s	1.439s	9.4x
large RPC response	reload	12.828s	1.313s	9.8x
metadata-heavy resume directory	resume	5.845s	0.535s	10.9x
metadata-heavy resume directory	reload	0.317s	0.265s	1.2x
mixed render workload	resume	3.997s	3.968s	1.0x
mixed render workload	reload	4.051s	3.797s	1.1x

The committed benchmark can be run with:

make bench-reload-resume          # GUI/xvfb, primary lane
make bench-reload-resume-batch    # batch, secondary lane
make bench-reload-resume-smoke    # cheap CI-friendly smoke

Validation

• make check
• reload/resume benchmark smoke in batch and GUI/xvfb
• reload/resume before/after benchmark run with synthetic fixtures