Correct repeated Transfer-Encoding and Content-Length response framing

Follow-up from the post-merge review of #23.

## Problem

The new HTTP/1 response pump uses `header_field`, which returns only the first field with a given name. Response framing therefore does not correctly handle repeated Transfer-Encoding or Content-Length fields.

Examples:

- `Transfer-Encoding: gzip` followed by `Transfer-Encoding: chunked` is validly terminated by chunked framing, but the proxy sees only the first field and treats the response as read-to-close.
- Conflicting Content-Length fields are not rejected.
- A malformed first Content-Length becomes indistinguishable from no framing and is treated as read-to-close.

On keep-alive connections this can consume subsequent responses as part of the preceding body, break request/response correspondence, or stall until the upstream closes.

## Location

- `src/http_proxy.rs`: `header_field`, `content_length`, `is_transfer_encoding_chunked`, and `response_body`
- Response pump introduced by #23.

## Acceptance criteria

- Combine all Transfer-Encoding field values in wire order and frame as chunked only when the final coding is `chunked`.
- Validate every Content-Length occurrence, including comma-list normalization where supported; reject conflicting or malformed values.
- Define and test behavior when Transfer-Encoding and Content-Length coexist in a response.
- Add response-side tests for:
- repeated TE ending in chunked;
- repeated TE not ending in chunked;
- identical repeated CL;
- conflicting CL;
- malformed CL;
- each case followed by another keep-alive response to verify correspondence.

Start in src/http_proxy.rs by reading header_field, content_length, is_transfer_encoding_chunked, and response_body, then trace the response pump introduced by #23. Add response-side coverage for repeated and malformed Transfer-Encoding and Content-Length cases, including each followed by a keep-alive response. Done means framing, rejection, and response correspondence match every acceptance criterion.

Written by the indexing model from the issue text.