Positive sub-microsecond socket timeouts disable the timeout on POSIX #220

On POSIX, a positive socket timeout shorter than one microsecond is converted to a zero timeval. SO_RCVTIMEO and SO_SNDTIMEO interpret zero as no timeout, so a small positive deadline can turn into an unbounded blocking operation.

Regression coverage first: test/regression/SocketTinyTimeout, in the existing regression layout. Connect two loopback sockets. Have the peer send one byte after 100 ms so the test remains bounded even when the timeout is disabled. A read with timeout 0.0000001 seconds should time out before the byte arrives. Repeat with 0.01 seconds as a control and wait for the sender to finish before closing sockets.

Confirmed on Linux x86_64 with Haxe 4.3.3:

tiny positive timeout expired: false
normal positive timeout expired: true

In src/hx/libs/std/Socket.cpp, socket_set_timeout truncates the fractional part to integer microseconds. The POSIX branch does not ensure that a positive request produces a nonzero timeval. The Windows branch already has the equivalent guard for positive sub-millisecond values.

Round a positive timeout up to the platform's minimum nonzero representation. Keep exactly zero and null as the existing no-timeout setting. Cover values below one microsecond, normal fractional values and the boundary near a full second, and avoid producing an invalid tv_usec of 1000000 through rounding.

The reproduction is Linux-only. This is separate from interrupted select/poll deadlines: the timeout is disabled before the receive operation even starts. No runtime implementation changes are included.