OOM when loading z-lab/gemma-4-26B-A4B-it-PARO

[zviratko]

I tried describing it in detail at jundot/omlx#1757
Not sure where the bug(s) lie but I think there are several and when they combine they are quite deadly :)

1. memory explosion upon load without release on failure
2. VLM->LLM fallback causing even more memory use
3. subsequent OOM when it doesn't fit
4. successful release when it does

I don't know if vision is supposed to work or not in this quant, but it works with gemma-4-26B-A4B-it-oQ8

[liang2kl]

Thanks for flagging the problems. I'll take a look.