Allow send capacity buffering beyond a full congestion window

[jeniksv]

tx_cap_factor is intended to allow applications to buffer additional stream data ahead of the congestion window. Currently, tx_cap is calculated as minimum of available congestion window and available flow control multiplied by tx_cap_factor. This means that once cwnd_available reaches zero, tx_cap also becomes zero regardless of tx_cap_factor. As a result, the application cannot pre buffer additional stream data while the congestion window is full and quiche needs to wait for more application data after ACKs make congestion window capacity available again.

This change keeps the existing cwnd_available() as the base capacity and adds an extra buffering allowance derived from the total congestion window:

base_cap =
    min(cwnd_available, flow_control_available)
    * min(tx_cap_factor, 1)

extra_cap =
    max((tx_cap_factor - 1) * cwnd - tx_buffered, 0)

tx_cap =
    min(base_cap + extra_cap, flow_control_available)

For values greater than one, the additional capacity no longer depends on cwnd_available being non-zero and quiche can keep stream data buffered even when the current congestion window is fully occupied.

Already buffered stream data is deducted only from the additional buffering allowance, preventing repeated tx_cap updates from granting the same extra capacity again.

The peer's connection-level flow-control limit remains an absolute upper bound and is not scaled by tx_cap_factor.