fix: bound PTY output memory growth

[htazq]

Supersedes #27.

Summary

• Keep the SSH channel output reader draining even when the downstream PTY output queue is full.
• Drop overflow at the SSH output boundary with a bounded queue instead of blocking on output_tx.send().await, which could allow russh/internal buffers to grow when frontend output stalls.
• Add ignored local stress tests for both direct PTY backpressure and the full WebSocket PTY stall path.
• Add a local Paramiko SSH stress server used by those ignored tests.
• Fix existing test blockers so the normal test suite can run cleanly: skip hard-coded SSH integration tests unless explicitly enabled, add missing xterm mock reset(), and preserve backward compatibility when tabToGroupMap is absent during terminal group serialization.
• Apply cargo fmt to the Rust workspace.

Root Cause

The previous main-branch implementation and #27 both focused on bounding downstream WebSocket/frontend output. Real stress testing showed that blocking the PTY/WebSocket reader does not reliably bound process memory. When the downstream path stalls, the SSH channel reader can stop draining channel.wait(), and output may still accumulate upstream inside russh/internal buffers.

This PR moves the memory boundary closer to the SSH channel reader: the reader keeps draining remote output and drops overflow before it can accumulate in the SSH stack.

Validation

• cargo fmt --check
• cargo check
• cargo test --lib — 77 passed, 7 ignored
• direct PTY stress: test_pty_output_memory_stays_bounded_when_output_is_not_consumed with local Paramiko SSH server, 60s, max private-memory growth about 3.3 MB
• WebSocket PTY stall stress: test_websocket_pty_memory_stays_bounded_when_frontend_stalls with local Paramiko SSH server, 60s, 4 output frames / 149,504 bytes, max private-memory growth about 3.8 MB
• pnpm exec tsc --noEmit
• pnpm test — 406 passed, 7 skipped
• pnpm lint — 0 errors, existing warnings remain
• pnpm build