docs: update parser launch examples

[mmangkad]

Summary

I noticed the Kimi and GPT-OSS launch docs were using parser names that the bundled SMG gateway does not accept. With the documented Kimi values, startup fails during SMG argument validation even after the TokenSpeed engine finishes warmup, so I updated the docs to use the accepted parser names: kimi_k25 / kimik2 for Kimi, and base for the GPT-OSS reasoning parser used by the existing CI recipe.

I also updated the server argument help text so it no longer lists gpt-oss as an example reasoning parser.

Currently accepted SMG parser choices

Reasoning parsers:

base
cohere_cmd
deepseek_r1
deepseek_v31
glm45
kimi
kimi_k25
kimi_thinking
minimax
nano_v3
none
qwen3
qwen3_thinking
step3

Tool-call parsers:

cohere
deepseek
deepseek31
deepseek32
deepseek_v4
glm45_moe
glm47_moe
json
kimik2
llama
minimax_m2
mistral
passthrough
pythonic
qwen
qwen_coder
qwen_xml
step3

Test Plan

I hit this startup failure when using the documented Kimi parser values:

[smg] __main__.py: error: argument --reasoning-parser: invalid choice: 'kimi_k2' (choose from 'base', 'cohere_cmd', 'deepseek_r1', 'deepseek_v31', 'glm45', 'kimi', 'kimi_k25', 'kimi_thinking', 'minimax', 'nano_v3', 'none', 'qwen3', 'qwen3_thinking', 'step3')
startup failed: smg subprocess exited with rc=2 during startup; see [smg] log lines above for the cause

After changing the launch command to use --reasoning-parser kimi_k25 and --tool-call-parser kimik2, the SMG parser validation issue is gone and the server can continue past gateway startup.