feat: add Together AI as a supported model provider

[Sayt-0]

Summary

Adds first-class support for Together AI as an OpenAI-compatible model provider via the alias map, following the existing fireworks/cerebras pattern.

Closes #3348

Definition of Done

Item	Status	Where
Alias entry together (openai, https://api.together.xyz/v1, TOGETHER_API_KEY)	done	pkg/model/provider/aliases.go
Unit test asserts the alias resolves and IsCatalogProvider("together") is true	done	TestCatalogAliases (together row), pkg/model/provider/aliases_test.go
Provider appears in the model catalog via CatalogProviders()	yes (has a BaseURL)	aliases.go
task test and task lint pass	yes (see Validation)
Example YAML config using a Together AI model	done	examples/together.yaml

Beyond the stated DoD, this also wires auto-detection, a default model, the JSON schema, and the provider docs, mirroring how Fireworks/Cerebras/DeepSeek were added.

Design decisions

Decision	Rationale
Added to openModelHostProviders (system-message coalescing)	Together fronts open-weight models (Llama, Qwen, DeepSeek, Kimi, GLM) whose chat templates may reject more than one leading system message. Consecutive system messages are coalesced, per #3344 / #3145 / #2327. First-party APIs with a fixed lineup (mistral, xai, deepseek) are deliberately excluded.
Listed in modelsDevAbsentProviders (example test)	models.dev catalogs Together under the provider id togetherai, not together, so the example-validation lookup is skipped — the same treatment as fireworks. No custom catalog-name mapping was added, keeping Together consistent with its sibling providers.
Auto-detection priority after cerebras, before amazon-bedrock	Keeps the open-model-host ordering consistent; precedence assertions updated accordingly.
Default model meta-llama/Llama-3.3-70B-Instruct-Turbo	Well-known general-purpose Together open model, present in the models.dev togetherai catalog.
Removed pkg/model/provider/openai/repro_issue3344_test.go (scratch repro file)	Its reusable TestShouldMergeConsecutiveMessages_Gating and an equivalent openai+base_url (vLLM) merge regression were moved into the properly-named system_message_merge_test.go, so no coverage is lost.

Tests

Test	Coverage
TestCatalogAliases (together)	Alias resolution and catalog membership.
TestOpenAIAliasProvider_EndToEndRequest (together)	Real HTTP request through the full stack against a local OpenAI-compatible server: asserts POST, Bearer TOGETHER_API_KEY, /chat/completions route, verbatim model, stream reassembly, and that consecutive system messages are coalesced into one.
TestOpenAIAliasProvider_LiveAPI (together)	Live API call, skipped unless TOGETHER_API_KEY is set, so the default run stays hermetic.
TestShouldMergeConsecutiveMessages_Gating (together)	Asserts coalescing is enabled for together.
auto_test.go cases	Detection, default model, precedence chain, integration with DefaultModels.
TestParseExamples	Validates examples/together.yaml.

Validation

Gate	Result
task build	passes
task lint (golangci-lint + custom cops)	0 issues / no offenses
go test ./...	touched packages pass; remaining failures are pre-existing SSRF/network tests unrelated to this change
CI on this PR	all checks green