RPC Binder: socket ENOMEM handling

(urgent to merge)

RPC Binder uses non-blocking sockets. This is so when client/server
pairs, when both writing to the same socket, if the socket would
block, they can back off to polling so that they can read commands
from the other side and recover.

When writing to one of these sockets, the vhost-vsock implementation
we are using returns ENOMEM in some cases.

In order to resolve this, we add an exponential packet size backoff
and time backoff.

There is the suggestion to use blocking sockets instead, since
it would engage direct reclaim that could potentially block. However,
I've not gone with this approach because:
- all RPC binder sockets are marked non-blocking globally. Setting
  the per-syscall flags on each operation all throughout the binder
  code base is a more invasive option. Specifically, it is hard
  to guarantee that we never hang indefinitely.
- the timeout needs to be configured with the global SO_SNDTIMEO
  socket option and can't be specified on sendmsg/recvmsg
  directly. This means retry would need to be added on every
  operation (if they now block, such as accept), or we would need
  2 syscalls for each of the operations here.

RPC_FLAKE_PRONE mode now simulates ENOMEM errors as well.

Future considerations:
- separate build of binderRpcTest with flake mode

Bug: 422574189
Test: binderRpcTest passes with FLAKE_MODE
Flag: EXEMPT bug fix

Change-Id: I5f2c99d0cb6760cca53ac4839d47daa9304a955a