Transactions perfomance improvements

[rugwirobaker]

Add a simple benchmark harness(to be improved)

• Allows me to run simple Writes, Reads, and Transactions benchmarks.
• Allows me to generate perf data and flamegraphs.
• Add corresponding documentation(DRAFT)

WAL changes

• Right now direct IO is a net negative until we can implement Async
  Direct io(ref: https://www.scylladb.com/2017/10/05/io-access-methods-scylla/).
• using fallocate(FALLOC_FL_KEEP_SIZE) on Linux.
• Simply WAL writes with buffered io(risk losing data on kernel
  crashes).
• Make the header read only therefore.

Reducing lock contention on the hot path.

• Remove cascading locks on WAL since it already had interior
  mutability.
• Replace RwLock with ArcSwap on Active memtable. More at
  docs/performance-profiling-2025-10-08.md.
• Since the header is read only we don't need to represent it in memory,
  therefore eliminating the need for the associated lock that was
  getting in the way of perfomance. See: docs/wal-header-design-flaw-2025-10-08.md

All in all we halved the number of futex calls per transaction