server: reduce synchronous PGAS polling latency

[vickiegpt]

Summary

• add adaptive server polling controls for the PGAS shared-memory path
• add explicit --pgas-record-accesses=false transport mode to skip unbounded controller occupation history while retaining validation, data transfer, synchronous responses, and protocol counters
• integrate the adaptive QEMU client from cxl: busy-poll PGAS responses on the hot path Zettai-US/qemu-cxl-type2#3
• record complete live devdax benchmark and counter evidence

Performance

• same-checkout original-wait control, true mediated: 31,963.106 ns/load at 128 MiB
• adaptive client plus lean server, true mediated: 5,009.636 ns/load at 128 MiB
• reduction: 84.3%
• the approximately 4.1 us installed-QEMU result is documented only as a non-mediated lower bound because server counters did not move

Correctness

• /dev/dax0.0 remained in devdax mode
• exact full command completed 115 numeric rows, zero zero-valued rows, exit 0
• no new invalid-opcode trap or lat_mem_rd segfault
• one isolated 100 ms response timeout is disclosed in the evidence

Verification

• Release build and CTest: 24/24 passed
• Debug build and CTest: 24/24 passed
• QEMU qemu-system-x86_64 target rebuilt successfully
• QEMU checkpatch: 0 errors, one generic MAINTAINERS warning for the new helper header