Fix: respect --max_seq_len for sliding window models with custom kv cache + sdpa

[rhn19]

Fixes #218

Changes

• _prepare_export_inputs: use max_dim = max_seq_len - 1 instead of
  min(max_seq_len, sliding_window) - 1 when ring cache is active
• export: bypass TorchExportableModuleWithHybridCache.__init__ to prevent
  StaticSlidingWindowLayer baking a <= sliding_window guard into the exported graph

Tested on

• google/gemma-3-1b-it with --max_seq_len 1024 --use_custom_kv_cache --use_custom_sdpa,
  prompt > 512 tokens