[Tuntii]

Moves viltrum to v1.15.4. Only VILTRUM_REF changes.

As promised in #1660, this batches several fixes into one bump rather than asking for a run per change:

• json-comp / json-h2c / json-tls: fixes the JSON encoder regression behind the dip in viltrum: update to 1.15.1 (buffers lent per event, static file cache) #1660's run (perf(jsonenc): write fields without attributes directly Tuntii/viltrum#121; plain JSON encode back to its pre-regression cost).
• TLS profiles (json-tls, static-h2, baseline-h2): AES-GCM now encrypts and authenticates in one pass, with Karatsuba GHASH (perf(tls): seal AES-GCM in one pass, GHASH with Karatsuba Tuntii/viltrum#124; +39% seal throughput on 16 KiB records on Zen 2).
• echo-ws / echo-ws-pipeline: WebSocket frames are written in one copy (perf(ws): append into spare capacity, small frames in one copy Tuntii/viltrum#125; -20% instructions per pipelined echo, -9% per single echo).

scripts/validate.sh viltrum: 98 passed, 0 failed (TLS verdict pass, tls_check pass).

Whenever there's a free slot, one run would be much appreciated. No rush, and thanks again for keeping these going.

[lalinsky]

I'm wondering if this shouldn't also be marked as "tuned" by following the same rules why e.g. cpeak is marked as tuned. As far as I can see, V has a JSON standard library. I think the rule needs a decision, because if it's really "language default" as we discussed in the dusty PR, then many entries will have to become tuned. Pretty much only C/C++ have the advantage of no JSON in the standard library, so frameworks have more freedom.