param: increase CQ read count to 16 for performance

This patch increases CQ read count to 16.

Experiments turned out that reading up to 16 (instead of 4) CQ entries
at once improves performance.

Signed-off-by: Michael Axtmann <axtmannm@amazon.com>
(cherry picked from commit 346be40)

include/nccl_ofi_param.h

@@ -129,7 +129,7 @@ OFI_NCCL_PARAM_INT(mr_key_size, "MR_KEY_SIZE", 2);
  * Maximum number of cq entries to read in a single call to
  * fi_cq_read.
  */
-OFI_NCCL_PARAM_INT(cq_read_count, "CQ_READ_COUNT", 4);
+OFI_NCCL_PARAM_INT(cq_read_count, "CQ_READ_COUNT", 16);
 
 /*
  * Protocol to use for send/recv operations.  Valid options are