docs(proposals): add What?/Why? for A2A task route capture parse comp…

[mkoushni]

Summary

try_capture_from_buffer in filters/src/agentic/a2a/mod.rs (the
non-streaming A2A task-route capture path) hex-decodes and attempts a
full serde_json::from_slice parse over the entire accumulated
response buffer on every on_response_body chunk, discarding the
result whenever the JSON is still incomplete. For a body delivered in
k chunks, this repeats full decode+parse work over an ever-growing
buffer on every chunk, producing O(n²) total work in response body
bytes rather than O(n) — see #353.

This PR adds the proposal's What?/Why? sections per the
docs/proposals/template.md convention. Per the template note, the
How? section is intentionally omitted here and will follow in a
separate PR once the proposal direction is agreed.

What's in this PR

• New proposal: docs/proposals/00353_a2a-task-route-capture-parse-complexity.md
• Describes bounding the per-chunk capture cost to O(chunk) via an
  incremental, string-aware brace/bracket depth scan, so the
  expensive hex-decode + JSON parse is only attempted once the buffer
  plausibly holds a complete, balanced JSON value.
• Explicitly scoped to the non-streaming capture path only — the
  SSE capture path (process_sse_response_chunk) already scans
  incrementally and is unaffected.
• Preserves the existing opportunistic, pre-EOS capture requirement
  (a route must be captured as soon as the JSON body is complete,
  without waiting for end_of_stream, since Pingora may not deliver
  a separate EOS callback after the final data chunk).
• Non-goals: no change to the hex-encoding scheme used for multibyte
  UTF-8 chunk-boundary safety, no change to the SSE scanner, and no
  overlap with fix(a2a): streaming methods with non-SSE response skip task route capture #352 (handled separately).

Test plan

• make lint passes (clippy, fmt-check, separator/filter-docs lints)
• No code changes in this PR — documentation only

Relates to #353