feat(batch-processor): add max_buffer_bytes to bound buffered/in-flight memory

[AlinsRan]

Description

Fixes #11244

Root cause

core.response.hold_body_chunk accumulated every response-body chunk into a Lua table ({chunk, n, bytes, ...}) and, when the size limit was reached or on eof, concatenated the whole table into one string with table.concat. For large bodies (the issue reports ~128K), this keeps both the list of individual chunks and the final concatenated string alive simultaneously, and each table.concat allocates a brand new string. The result is excessive allocation/retention that shows up as memory growth when loggers hold large response bodies.

Fix

Rewrite the accumulation to use a LuaJIT string.buffer (require("string.buffer")), reviving the approach from the abandoned #12035 and adapting it to the current hold_body_chunk shape:

• chunks are appended with buffer:put(chunk) instead of being stored in a growing Lua table;
• the final body is produced with buffer:get() (full body) or buffer:get(max_resp_body_bytes) (truncated), which returns exactly the first max_resp_body_bytes bytes and avoids materializing an intermediate concatenated copy before truncation;
• the wrapper table is kept only for the bytes counter and the done flag so the buffer object can carry per-buffer_key state.

All existing semantics are preserved:

• the hold_the_copy flag (still flushes arg[1] on non-final chunks when false);
• max_resp_body_bytes truncation (returns exactly max_resp_body_bytes);
• the done flag (later chunks and eof return nil after a truncated body was already returned);
• the single-chunk-that-is-also-eof path;
• per-body_buffer_key / ctx._plugin_name storage.

Test

Adds t/core/response-hold-body.t covering:

1. multi-chunk accumulation of a 128K body (full body returned at eof);
2. truncation at max_resp_body_bytes across multiple chunks;
3. the done flag — later chunks and eof return nil after truncation;
4. a single chunk that is also eof, truncated to max_resp_body_bytes;
5. small-body single-chunk passthrough (mirrors the existing t/core/response.t TEST 8 mock pattern).

Verification

• luacheck apisix/core/response.lua passes (0 warnings / 0 errors).
• All inline Lua blocks in the new .t parse cleanly.
• The rewritten function was exercised end-to-end via a standalone harness that drives the same ngx.arg[1]/[2] chunk/eof protocol; all six scenarios above pass and return byte-identical bodies (e.g. truncation returns exactly 131072 bytes).
• A memory micro-benchmark inside test-nginx is not reliable, so the rationale is documented above; the string.buffer path avoids the simultaneous chunk-table + concatenated-string retention that caused the growth.

Note: the full prove test-nginx run was not executed in my local environment because the available OpenResty build (1.21.4.4) does not support the quic/http3 listen directives that t/APISIX.pm emits unconditionally on master; this is an environment limitation, not a test failure. CI runs against a QUIC-capable OpenResty.

Checklist

• I have explained the need for this PR and the problem it solves
• I have explained the changes or the new features added to this PR
• I have added tests corresponding to this change
• I have updated the documentation to reflect this change
• I have verified that this change is backward compatible