For the dynamic I/O scheduler, make the TRIM stuff also count against

For the dynamic I/O scheduler, make the TRIM stuff also count against
read bias so we do reads in preference to TRIMs. This helps a lot when
many trims are delivered at once from the upper layers as they tend to
delay READs due to priority inversion in the code today.

The non iosched case will be fixed when the trim comibing changes
needed for nvme come in later this year.

Sponsored by: Netflix