fix: add explicit size caps for config reads, cached notifications, and LSP errors

[bug-ops]

Summary

Several inputs were previously bounded only by an outer transport or protocol limit, or not bounded at all — each a defense-in-depth gap found during a security audit (CWE-400).

• ServerConfig::load_from now reads the config file through a bounded Read::take instead of an unbounded read_to_string, rejecting files over 8 MiB. A metadata() size pre-check alone is insufficient since it reports 0 for character devices and FIFOs (e.g. /dev/zero).
• rename_symbol's new_name and get_completions's trigger MCP params are now length-capped, matching the existing cap on workspace_symbol_search's query.
• NotificationCache::store_log/store_message/store_diagnostics now truncate cached text and bound each diagnostics entry's total serialized size, instead of capping entry count only. Diagnostics truncation prefers keeping higher-severity entries (largest-fitting-prefix via binary search, not a lossy positional halve) and logs a warning when content is dropped.
• LspClient now forwards a bounded error message to MCP callers instead of the full, unbounded LSP server error text.

Fixes #309, #311, #313

Test plan

• cargo +nightly fmt --all -- --check
• cargo clippy --all-targets --all-features --workspace -- -D warnings
• cargo nextest run --workspace --all-features --lib --bins (691 passed, 1 skipped)
• RUSTDOCFLAGS="-D warnings" cargo doc --no-deps --all-features
• CHANGELOG.md updated under [Unreleased] > Security