Remove allow_refill flag arguement from _oni_read_buffer

[jonnew]

• _oni_read_buffer was called twice in oni_read_frame: once for the frame header with allow_refill=1, and once for the payload with allow_refill=0. The flag was necessary because a second refill during the payload read would change ctx->shared_rbuf, breaking the single-buffer invariant that frames depend on.

• The flag is unnecessary because _oni_read_buffer already refills to block_read_size bytes, and block_read_size >= max_read_frame_size >= max_data_size. After the header read the buffer is always guaranteed to hold the full payload, so a second refill can never occur. The payload read is now inlined in oni_read_frame.

• The ONI_ENOREADDEV guard for max_read_frame_size == 0 is moved into _oni_read_buffer, which already uses that field as a refill threshold, making the function self-contained and allowing the NULL check on shared_rbuf within _oni_read_buffer to be removed.

[aacuevas]

Since we are moving towards unnecessary parameters, and since now _oni_read_buffer is only called once with a size parameter of always ONI_RFRAMEHEADERSZ, shouldn't we take the opportunity to remove the size parameter altogether?

Maybe just rename it to something like int _oni_ensure_buffer(oni_ctx ctx) that performs the buffer allocation if necessary and adjust the pointers appropriately, and have oni_read_frame read the header through these pointers, as it already does for the payload.

Right now it reads the payload through the _oni_read_buffer returned pointer but then uses the context pointers for the payload, which seems odd.

[jonnew]

This should be addressed in the latest commit.