net: macb: manage BNA error and prevent RX lockup on GEM

[yseraf]

This patch addresses a controller freeze occurring under high incoming traffic loads on GEM-based platforms like the ATSAMA5D41. When the DMA exhausted all available buffers, it triggered a Buffer Not Available (BNA) condition in the Receive Status Register (RSR).

On this hardware, clearing the BNA flag without first resolving the underlying resource shortage caused the controller to enter a permanent blocked state. The recovery logic introduced here forces a slot release by incrementing the consumer index (rx_tail), acknowledges the error in RSR, and performs a descriptor refill before allowing NAPI to continue. To isolate this critical repair phase from asynchronous bus errors, the HRESP interrupt is transiently masked. This GEM-specific implementation is separated from the legacy MACB polling to avoid regressions on older architectures.

[yseraf]

The bug and this patch can be easily tested using Packet Sender application with its UDP Traffic Generator utils :

Without patch, the Ethernet controller gets stuck almost instantly (in few dozen of seconds).
With the patch, the Ethernet controller keep working even under heavy load for hours.

Note: the standard Ethernet controller behaviour and performance like bandwidth is not affected by the patch, it just makes the ATSAMA5Dxx a true network compatible CPU.