fixup: cpu: x64: improve tbb bnorm realtime inference performance

Realtime inference with HT on could use up to 8 threads.

src/cpu/x64/jit_uni_tbb_batch_normalization.cpp

@@ -2169,7 +2169,7 @@ struct driver_t : public c_compatible {
             dim_t total_size = size_src_dst + size_stats_ss_tensors;
 
             // Try to create at least nthr_ chunks for realtime inference
-            const int n_chunks_min = nthr_ <= 4 ? nstl::min(4, nthr_) : 1;
+            const int n_chunks_min = nthr_ <= 8 ? nthr_ : 1;
             const size_t l2_per_core = platform::get_per_core_cache_size(2);
             dim_t n_chunks
                     = nstl::max<dim_t>(n_chunks_min, total_size / l2_per_core);