docs: publish v2.2 as latest (#2111)

docs: publish v2.2 as latest (#2111)

docs: quote placeholders in shell examples so the blocks paste safely An unquoted <placeholder> inside a shell-tagged fence is redirection from a file of that name, so a copied command aborts (exit 1), fails to parse (exit 2), or — where it sits mid-block — is silently skipped while the block still reports exit 0. Quote the placeholder, which keeps it visible as a placeholder and passes it through verbatim. Usage synopses are left alone: where a fence carries [...] optional-argument notation, or a placeholder fills a command slot, the <> is grammar for the command's shape rather than a command to paste. Fixes #2196

docs: publish v2.2 as latest (#2111)

docs(models): paged attention is requested automatically, not unconditionally enabled The local-model paged-attention note collapsed three branches into one absolute sentence. `MistralLlama::paged_attention_config` returns `None` when the device is CPU or when the build is not `all(feature = "cuda", target_family = "unix")`, and the engine downgrades to dense attention again for architectures with no paged kernel (the Multi-head Latent Attention GGUFs) — so a CUDA+Unix build is necessary, not sufficient. From v2.2.0 (spiceai/spiceai#11990) the note is also wrong to say no configuration is needed: `FileModelParams::paged_attention` accepts `auto` (default) or `disabled`, and `disabled` forces dense attention. It is a `from: file:` parameter only — `HuggingFaceModelParams` has no such field and `from_hf` calls `paged_attention_config` unconditionally. Scoped accordingly: the branch distinction lands in v2.0.x-v2.3.x and vNext; the escape hatch only in v2.2.x, v2.3.x and vNext.

docs: publish v2.2 as latest (#2111)

docs(huggingface): the v2.2.x model token is huggingface_token, and bare token never worked

docs: publish v2.2 as latest (#2111)