Add Anthropic count_tokens endpoint

[iLoveBug]

Summary

• add the Anthropic-compatible POST /v1/messages/count_tokens endpoint
• reuse the /v1/messages parser, model-family prompt renderer, and loaded model tokenizer without queueing generation
• add a reusable integration test that compares count_tokens with total Anthropic input usage, including cache fields

Test plan

• make -B ds4-server ds4_test
• ./ds4_test --server
• DeepSeek V4 Flash IQ2_XXS/Q2_K integration test with ctx=4096: 344 input tokens
• DeepSeek V4 Flash IQ2_XXS/Q2_K integration test with ctx=32768: 344 input tokens