perf(ctx): cache parsed request body to avoid repeated decode for post_arg.*

[AlinsRan]

Summary

When multiple post_arg.* variables are evaluated in a single request
(e.g. vars expressions in route matching or plugin-level vars conditions
in traffic-split / fault-injection), each access independently
triggers a full body read + json.decode cycle. For large request
bodies at high QPS, this causes unnecessary repeated CPU and memory
overhead.

Changes

apisix/core/ctx.lua

Split get_parsed_request_body into two layers:

• _get_parsed_request_body(ctx) — pure parsing logic, no side effects
• get_parsed_request_body(ctx) — wrapper that caches the decoded table in ctx._post_arg_request_body for the lifetime of the request

On the first access, the decoded table is stored in ctx._post_arg_request_body.
Subsequent accesses to different post_arg.* keys within the same request
reuse the cached result, so json.decode runs at most once per request.

Errors are intentionally not cached: a failed parse leaves the cache
empty so that a later access (e.g. after a plugin sets a valid body) can
still succeed.

apisix/patch.lua

Patch ngx.req.set_body_data to automatically invalidate both cache layers
when the request body is rewritten mid-request:

• Clears api_ctx._post_arg_request_body (body-level cache)
• Removes all post_arg.* entries from api_ctx.var._cache (variable-level cache)

The invalidation is skipped entirely when _post_arg_request_body is nil,
avoiding any overhead on requests that never access post_arg.* variables.

Tests

• TEST 3/4 (t/core/ctx3.t): verifies all fields are read correctly and
  that reuse parsed request body from ctx cache appears exactly twice when
  three different post_arg.* keys are accessed in one request.
• TEST 5/6 (t/core/ctx3.t): verifies that after ngx.req.set_body_data(),
  both the same key and a different key reflect the new body content.