fix(server): honor prompt cache byte limit

[agisilaos]

Summary

mlx_lm.server parsed --prompt-cache-bytes, but the server-created LRUPromptCache was only initialized with prompt_cache_size. That left the LRU cache itself with its default effectively-unbounded byte limit, so sequential serving could retain prompt cache entries beyond the configured byte budget.

This wires prompt_cache_bytes into LRUPromptCache(max_bytes=...) when the server constructs the prompt cache, keeping byte-limit enforcement centralized for all cache insertion paths.

Changes

• Add make_lru_prompt_cache() to construct the server prompt cache with both sequence and byte limits.
• Use the helper from run() so --prompt-cache-bytes is honored by the LRU cache itself.
• Add a regression test showing a server-created prompt cache evicts entries according to the configured byte budget.

Alternatives considered

• Adding another trim call only in the sequential serve path. This was rejected because passing the byte limit into LRUPromptCache keeps enforcement centralized and applies consistently to every insertion path.

Notes

• Fixes the server prompt-cache byte-limit enforcement gap related to mlx_lm.server causes macOS kernel panic (IOGPUMemory crash) due to unbounded memory growth #883.
• Related to mlx_lm.server crashes on Metal OOM instead of returning an HTTP error #854 and generate() crashes on Metal OOM instead of recovering gracefully #1015, which track broader Metal OOM recovery behavior. This PR reduces one server-side OOM trigger but does not attempt to make arbitrary Metal OOMs recoverable.
• Overlaps with fix: honor --prompt-cache-bytes in sequential serve mode #1118.

Testing

• python3 -m unittest tests.test_server.TestLRUPromptCache
• black --check mlx_lm/server.py tests/test_server.py
• isort --profile=black --check-only mlx_lm/server.py tests/test_server.py