[TransferEngine] Add FlagCX support (#3522)

* flagcx support

* use flagcx p2p transport

* [Transport] FlagCX: batch slices per target into one multi-rail xfer

ioWorker now drains the whole io_queue and coalesces slices by (target,opcode)
into a single multi-iov WriteVector/ReadVector instead of one blocking
WriteVectorSync per block. With the FlagCX multi-rail engine this keeps every
NIC pool full (deep pipeline) instead of one block-at-a-time, roughly doubling
single-process throughput.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

* Enable FlagCX protocol in Python TE init

* Flagcx lat impr (#2622)

* implement direct submit

* move transfer status polling to getTransferStatus

* add lightweight completion polling thread

* [TransferEngine] Remove missing FlagCX demo target

* [TransferEngine] Apply minor FlagCX transport fixes

* [TransferEngine] Format FlagCX transport changes

---------

Co-authored-by: zhangzuoyuan <zhangzuo21@mails.tsinghua.edu.cn>
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: yuechen-sys <yuechen.bupt@gmail.com>
Co-authored-by: mikethegoblin <46526613+mikethegoblin@users.noreply.github.com>