fix(vllm): floor max cached blocks at 1 for sub-block-size budgets (#343)

[SuperMarioYL]

Summary

Fixes #343.

_get_max_cached_blocks (kvcached/integration/vllm/patches.py) converts the
unified KVCACHED_MAX_CACHED_TOKENS config into a block budget with:

return MAX_CACHED_TOKENS // block_size

When KVCACHED_MAX_CACHED_TOKENS is a small positive value below one KV
block (e.g. 8 with a 16-token block_size), integer division floors to 0.
That is the exact value the == 0 "disabled — evict on free" sentinel
produces, so a user who asked for a small non-zero cache gets prefix caching
silently turned off instead of a small cap.

Fix

Handle the sentinels explicitly and floor the positive branch at one block:

if MAX_CACHED_TOKENS < 0:
    return -1          # unlimited (unchanged)
if MAX_CACHED_TOKENS == 0:
    return 0           # disabled  (unchanged)
max_cached_blocks = MAX_CACHED_TOKENS // block_size
if max_cached_blocks == 0:
    logger.warning(...)  # no longer silent
    return 1
return max_cached_blocks

• A positive-but-sub-block budget now keeps prefix caching enabled at the
  smallest viable unit (1 block), which the consumer at
  ElasticBlockPool (max_cached_blocks >= 0 and len(evictable) > max_cached_blocks)
  treats as a valid cap of one evictable block.
• The clamp is no longer silent — a warning reports the effective
  granularity and points at KVCACHED_MAX_CACHED_TOKENS=0 for explicit disable.
• < 0 (unlimited), == 0 (disabled), and >= block_size budgets are
  byte-for-byte unchanged; the only behavioral change is the previously
  broken 0 < tokens < block_size case.

Tests

Adds tests/test_get_max_cached_blocks.py — GPU-free (no vmm_ops, GPU, or
installed vLLM), mirroring the existing tests/test_make_cache_key.py. Covers
the -1 / 0 / sub-block / exact-multiple / non-multiple / large boundaries, the
#343 regression (0 < tokens < block_size never collapses to the disabled
sentinel), and the warn-on-clamp / no-warn-on-full-block behavior.

$ pytest tests/test_get_max_cached_blocks.py -q
12 passed