net: macb: Fix RX ring refill issue after link up

[pelwell]

[ Patch created by @xiaoleiwang123456 ]

When the link goes down, the receive pipeline is cleared but the RX queue pointer register remains unchanged. This causes a critical issue where some RX ring entries may contain invalid addresses after link up.

The problem manifests when:

1. An entry with an invalid address prevents rx_tail from advancing (gem_rx() checks rxused as 0)
2. No gap exists between rx_prepared_head and rx_tail from previous refill operations
3. gem_rx_refill() cannot refill entries because CIRC_SPACE() always returns 0
4. RSR register reports exception with value 0x7
5. The hardware detected that the buffer was unavailable. The DMA will reread the pointer each time it receives an end-of-frame signal until a valid pointer is found, but the software cannot fill it.

This regression was introduced by commit 99537d5 which moved mog_init_rings() from macb_mac_link_up() to macb_open(), leaving the RX ring in an inconsistent state after link transitions.

Fix this by:

• Reinitializing RX rings on link up via mog_init_rings()
• Adding full_refill atomic counter to force complete ring refill
• Ensuring all entries are traversed and refilled after link up
• Resetting rx_tail and rx_prepared_head to 0

Fixes: 99537d5 ("net: macb: Relocate mog_init_rings() callback from macb_mac_link_up() to macb_open()")
Closes: #7200 (comment)