fix(completion): respect configured OpenAI prompt templates

[chyroc]

Summary

• Respect a non-empty prompt_template for openai/completion and
  openai/legacy_completion.
• Preserve the existing generic FIM fallback and prompt / suffix transport
  when no template is configured.
• Add request-level coverage for the serialized /v1/completions payload.

Context

Related to #4523, for the OpenAI-compatible completion path only. This change
does not modify vllm/completion and does not claim to resolve the issue's
reported vLLM behavior.

PR #3341 intentionally gave the OpenAI aliases Tabby's generic
{prefix}<|FIM|>{suffix} default. This change keeps that default when
prompt_template is absent or empty, but lets an explicit model-specific
template override it. An explicitly configured template containing Tabby's
internal <|FIM|> marker also retains the existing split prompt / suffix
transport. deepseek/completion, mistral/completion, and
vllm/completion keep their current behavior.

Test plan

CARGO_NET_OFFLINE=true cargo test --locked -p http-api-bindings --lib
cargo fmt --all -- --check
CARGO_NET_OFFLINE=true cargo clippy --locked -p http-api-bindings \
  --all-targets --no-deps -- -D warnings
git diff --check

Results: 6 affected-crate tests passed, 0 failed, and 3 existing network
integration tests were ignored. The regression starts a loopback HTTP server
and checks the request path and complete JSON body for a configured Qwen-style
template, the default fallback, and an explicit internal FIM template.

The workspace-wide test command was also attempted, but this shallow checkout
does not contain the recursive llama.cpp submodule, so the unchanged
llama-cpp-server build script stopped before workspace tests ran.