Issue #1230: transport/ofi: do not constrain inject_size hint to sizeof(long double)

[bcmIntc]

Summary

Addresses Issue #1230.

• query_for_fabric previously initialized shmem_transport_ofi_max_buffered_send = sizeof(long double) before calling fi_getinfo, which passed that value as
  tx_attr.inject_size — telling the provider it must support at least 16-byte inject. This caused fi_getinfo to filter out providers that advertise a smaller
  natural inject size.
• This change sets the hint to 0 (no minimum constraint) so the provider reports its natural inject_size. The returned value is immediately adopted after
  fi_getinfo, and the existing assert (inject_size >= max_buffered_send) trivially holds since both are 0 at that point.
• On providers where the natural inject size is smaller than 16 bytes, the previous behavior could silently fail to select an otherwise valid endpoint, or cause
  fi_getinfo to return an inflated inject_size that doesn't reflect actual NIC capability.

Test plan

• Build and make check with CXI provider
• Build and make check with verbs provider
• Confirm shmem_transport_ofi_max_buffered_send value logged at startup reflects the provider's natural inject size
• Verify no regression in small-message latency (inject path still engaged for messages <= provider's natural inject size)