Build the EnforceDepthLimit performance requests without recursion, and stop tagging the performance suite.

The three EnforceDepthLimit tests in binary_recursion_limit_test.cc built a 20000-deep (10000 for MapStringKey) message object and called SerializeAsString() on it. ByteSizeLong(), the serializer and the destructor each recurse once per nesting level, which overflows the 8 MB stack under the sanitizers' larger frames in the test binary itself, before any testee is involved; that is the only reason every `<x>_performance_test` was tagged noasan/nomsan/notsan/noubsan/nodiorite by conformance_test().

The payloads are now built directly on the wire with the binary_wireformat.h helpers: each nesting level is the bytes before the level below it (the level's own fields and the tag of the field holding the next level) and the bytes after it (a group's end tag), and Nested() computes the sizes inside out, then appends the bytes outside in, with the length prefixes in between. That is O(total bytes) time, O(depth) heap and constant stack. The builders live in their own library, recursion_limit_payloads.{h,cc} (DeepMapPayload(), DeepMapStringKeyPayload(), DeepMessageSetPayload()), so that the suite file holds only the three tests. The bytes are identical to what the message objects serialized: recursion_limit_payloads_test builds all three payloads both ways at depths 1, 3 and 50 (two-byte length prefixes) and a depth-0 message set, and compares them, and the checked-in request goldens of migration:request_golden_test are unchanged.

With that fixed, conformance_test() has no reason left to tag `<x>_performance_test` differently from the other suites' tests: `_PERFORMANCE_TEST_TAGS` and the performance special case go away, every test gets exactly the caller's tags plus `conformance`, and the bzl_tests now assert that `<x>_performance_test` carries `conformance` but not `noasan` (as a representative of the dropped tag set), like the binary test. All 14 `<x>_performance_test` targets in the chain pass under asan, tsan and ubsan, and under msan except java and java_lite, whose callers tag them nomsan (Java doesn't support MSan); all 14 also pass on ARM (--cpu=arm), no slower than on x86. merged_runner_test and migration:request_golden_test drop their sanitizer and nodiorite tags too: the legacy binary suite lost its recursive --performance requests when the binary performance tests moved to gtest (binary_recursion_limit_test.cc), so nothing the merged runner links builds a deep payload recursively any more, and both pass under asan, msan, tsan and ubsan and on ARM.

PiperOrigin-RevId: 984539805