server: ensure prompt caching for SWA models

[shipped-it]

Overview

When using --swa-full with Gemma 4 and other SWA models, the server can sometimes incorrectly force full prompt re-processing on every request, even though --swa-full allocates a full-size SWA cache where tokens are not pruned.

Fixed by:

• Setting pos_min_thold = 0 when swa_full is enabled (any cached position is useful)
• Skipping checkpoint restoration logic when swa_full is enabled (unnecessary since cache is full-size)

In short, this makes inference faster for SWA models when using prompt caching.

Closes #21468

Additional information

I was trying to debug a prompt caching issue: slot update_slots: id 0 | task 76 | forcing full prompt re-processing due to lack of cache data (likely due to SWA or hybrid/recurrent memory, see #13194 (comment))

1. I've read those issues and other PRs (to no avail):

• Eval bug: Context-shift with gemma 4 #21379
• kv-cache : add SWA support #13194
• server : fix SWA condition for full context reprocess #14163

2. But I could not find a solution to my problem, so I traced the error message to tools/server/server-context.cpp
3. Then found pos_min_thold that determines the minimum position that must be present in cache for restoration to work. It was always calculated as pos_next - n_swa, which for Gemma 4 means pos_next - 1024. I believe this is incorrect for swa_full, as with full-size cache, ANY cached position should be valid.
4. There was also a condition checking if checkpoint restoration was needed, but it didn't account for swa_full. With swa_full, the cache is never pruned so checkpoint restoration is unnecessary.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: NO

[elektricM]

Looks good to me, also fixed my issue on swa models on my amd card.