[performance-profiler] Reduce diskqueue reader_loop frame metadata read overhead in nextFrame

## Hot Path
`(*readerLoop).nextFrame` in `libbeat/publisher/queue/diskqueue/reader_loop.go` (`L191-L246`) repeatedly parses frame metadata with multiple small reads (`L201`, `L229`, `L240`) and uses `Read` for payload (`L222`).

## Profiling Data
**Before:**
```text
go test ./libbeat/publisher/queue/diskqueue -run=^$ -bench=BenchmarkSync1k -benchmem -count=5 -cpuprofile=/tmp/gh-aw/agent/dq_cpu.prof -memprofile=/tmp/gh-aw/agent/dq_mem.prof
BenchmarkSync1k-4 99 12022922 ns/op 3466193 B/op 33291 allocs/op
BenchmarkSync1k-4 99 12762895 ns/op 3462574 B/op 32998 allocs/op
BenchmarkSync1k-4 96 12038660 ns/op 3465094 B/op 33219 allocs/op
BenchmarkSync1k-4 99 11884420 ns/op 3466151 B/op 33329 allocs/op
BenchmarkSync1k-4 99 12225757 ns/op 3465529 B/op 33288 allocs/op
```

CPU profile top (before):
```text
go tool pprof -top /tmp/gh-aw/agent/dq_cpu.prof
... total samples = 23.45s
internal/runtime/syscall.Syscall6 11.58s (49.38%)
```

## Proposed Change
Replace per-field `binary.Read` calls with `io.ReadFull` + `binary.LittleEndian.Uint32` and read footer in one 8-byte read.

```diff
diff --git a/libbeat/publisher/queue/diskqueue/reader_loop.go b/libbeat/publisher/queue/diskqueue/reader_loop.go
@@
-var frameLength uint32
-err := binary.Read(reader, binary.LittleEndian, &frameLength)
+var header [4]byte
+_, err := io.ReadFull(reader, header[:])
+frameLength := binary.LittleEndian.Uint32(header[:])
@@
-_, err = reader.Read(bytes)
+_, err = io.ReadFull(reader, bytes)
@@
-var checksum uint32
-err = binary.Read(reader, binary.LittleEndian, &checksum)
+var footer [8]byte
+_, err = io.ReadFull(reader, footer[:])
+checksum := binary.LittleEndian.Uint32(footer[0:4])
@@
-var duplicateLength uint32
-err = binary.Read(reader, binary.LittleEndian, &duplicateLength)
+duplicateLength := binary.LittleEndian.Uint32(footer[4:8])
```

## Results
**After (same command):**
```text
go test ./libbeat/publisher/queue/diskqueue -run=^$ -bench=BenchmarkSync1k -benchmem -count=5 -cpuprofile=/tmp/gh-aw/agent/dq_cpu_after.prof -memprofile=/tmp/gh-aw/agent/dq_mem_after.prof
BenchmarkSync1k-4 114 10441470 ns/op 3468320 B/op 31744 allocs/op
BenchmarkSync1k-4 115 11351108 ns/op 3466468 B/op 31614 allocs/op
BenchmarkSync1k-4 106 10575430 ns/op 3466453 B/op 31720 allocs/op
BenchmarkSync1k-4 115 10384787 ns/op 3467907 B/op 31805 allocs/op
BenchmarkSync1k-4 100 10794191 ns/op 3467981 B/op 31906 allocs/op
```

Averages (5 runs):
- Time: `12,186,930.8 -> 10,709,397.2 ns/op` (**12.12% faster**)
- Memory: `3,465,108.2 -> 3,467,425.8 B/op` (roughly flat)
- Allocs: `33,225.0 -> 31,757.8 allocs/op` (**4.42% fewer allocs**)

**Improvement:** 12.12% latency reduction on a diskqueue hot path benchmark.

## Verification
- Tests run: `go test ./libbeat/publisher/queue/diskqueue -run 'Test.*' -count=1` → `ok`
- Behavior preservation: checksum validation, duplicate-length validation, and decode flow are unchanged; only byte-read/parsing mechanics changed.

## Evidence
Commands executed:
- Baseline benchmark/profile:
- `go test ./libbeat/publisher/queue/diskqueue -run=^$ -bench=BenchmarkSync1k -benchmem -count=5 -cpuprofile=/tmp/gh-aw/agent/dq_cpu.prof -memprofile=/tmp/gh-aw/agent/dq_mem.prof`
- Post-change benchmark/profile (same command):
- `go test ./libbeat/publisher/queue/diskqueue -run=^$ -bench=BenchmarkSync1k -benchmem -count=5 -cpuprofile=/tmp/gh-aw/agent/dq_cpu_after.prof -memprofile=/tmp/gh-aw/agent/dq_mem_after.prof`
- Verification tests:
- `go test ./libbeat/publisher/queue/diskqueue -run 'Test.*' -count=1`

This appears distinct from existing open issue #49519 (buffer reuse in `serialize.go`), as this targets metadata parsing in `reader_loop.go`.

---
[What is this?](https://ela.st/github-ai-tools) | [From workflow: Performance Profiler](https://github.com/elastic/beats/actions/runs/23494539632)

Give us feedback! React with 🚀 if perfect, 👍 if helpful, 👎 if not.
> - [x] expires on Mar 31, 2026, 2:42 PM UTC

Start in libbeat/publisher/queue/diskqueue/reader_loop.go, especially readerLoop.nextFrame at lines 191-246, and review the existing frame layout and validation flow. Run the BenchmarkSync1k benchmark and the diskqueue tests before making changes. Done means preserving checksum, duplicate-length, and decode behavior while matching the reported performance improvement.

Written by the indexing model from the issue text.