Fix tactic cache keying, plugin flush, and rebase regressions
- Restore setL2CacheBytes/mL2CacheBytes and the FORCE_CHUNKED doc text dropped by the rebase.
- Key the in-process profile map by bits/group size/zero points/GEMV support so kernels with
different quantization do not share tactics or persist them under the wrong cache key.
- Stage lazily profiled buckets right after profiling instead of in kernel destructors, so any
flush point sees them regardless of teardown order.
- Flush from ~CudaEp: the plugin build does not compile cuda_execution_provider.cc.
- Log flush failures at EP teardown.