Improve nvram performance

[zestrada]

Our nvram accesses could be faster, and this might be a performance bottleneck for systems that make frequent accesses to that subsystem.

Two thoughts on this:

1. We make hypercalls in our nvram_ functions for monitoring the user may or may not be interested in. We should the hypercalls themselves optional so we don't call out to the hypervisor if we aren't interested in logging those events.
2. We have a tmpfs-based key/value store. That involves a lot of syscalls to access, maybe we should consider an alternative.