server: add --reasoning and --alias CLI options - #1017

Motivation

Clients that expose a model picker can inject thinking defaults that conflict
with the mode a user selected. This adds server-side defaults plus a way to
advertise an instance under a custom model id.

Changes

• --reasoning on|off / --thinking on|off — default thinking state when a
  request omits one. The two are synonyms; the error message reports the flag
  the user actually typed.
• --no-thinking — shortcut for --reasoning off.
• --reasoning-effort none|minimal|low|medium|high|xhigh|max — default effort.
• --alias NAME — advertise an extra model id in the catalogue so a picker can
  target this instance by name. The id is also served by GET /v1/models/NAME.

Alias suffixes select the default thinking mode: -chat/-nothink disable it,
-reasoner enables it. The match is on the suffix rather than a substring, so a
local quant name such as local-glm-5.3-flash-q2 stays unaffected while a
...-chat alias still opts out.

Turning the default off must not disable thinking permanently

think_mode_from_enabled() returns NONE whenever the effort is NONE, so an
earlier revision that set default_think_mode = NONE on --reasoning off
pinned every request to non-thinking. A client that opted back in through a
-reasoner alias or an explicit thinking: true was silently ignored.

Off now flips only the boolean and keeps the configured effort:

--reasoning off                  → enabled=false, effort=HIGH
  request without thinking field → NONE
  request with thinking:true     → HIGH   (was NONE before this fix)

--reasoning off --reasoning-effort max
  request with thinking:true     → MAX    (was NONE before this fix)

--reasoning-effort none          → deliberate opt-out, stays NONE

Catalogue bug fix

send_models() picked the advertised ids with a family test, but GLM-5.3 and
GLM-5.2 share DS4_MODEL_FAMILY_GLM_DSA. A GLM-5.3 engine therefore advertised
glm-5.2, glm-5.2-chat and glm-5.2-reasoner, misreporting the loaded model
to a picker. The ids are now selected by variant, and a single list drives both
the catalogue and the alias-deduplication check.

Testing

Machine: Apple M4, macOS 26.4.1, 16 GB. Backend: Metal (default build).

make ds4_test
./ds4_test --server
→ server: OK

--server covers the request parsing, chat rendering and thinking-control paths
this change touches. It needs no model file. The model-backed suites
(--logprob-vectors, --long-context, ...) were not run because this machine
has no GGUF checked out.

New tests:

• test_server_reasoning_options — flag parsing, the off/on effort combinations
  above, plus negative alias-suffix cases (my-chat-model,
  local-chat-assistant must not disable thinking).
• test_model_catalog_matches_loaded_variant — the three variant catalogues,
  and every advertised id is accepted by server_model_alias_known().
• test_send_models_json_shape — drives the real /v1/models body over a
  socketpair and checks the JSON shape, plus that a built-in alias is not
  advertised twice.

No inference-backend code is touched, so no speed regression is expected.