md: Get rid of the pbuf zone

[markj]

Summary

The zone is used to provide KVA for mapping BIOs so that we can pass
buffers to VOP_READ and VOP_WRITE. Its use doesn't really make sense:
it is limited globally to nswbuf/10 allocations, so this might not work
well if a large number of vnode MDs are in use. Furthermore, all I/O to
a MD device is handled with a dedicated thread, so we might as well just
have one reserved block per softc.

I suspect we should go further and add VOPs which enable reading and
writing with unmapped buffers. But this step lets us get rid of some
unncessary pbuf pre-allocations, which is my current goal.

[kib]

The use of dedicated (sub)zone is reasonable for KVA-starved arches, where we do not want to waste 1MB of KVA per md. Instead KVA use was shared with other consumers.

[markj]

In D40215#915694, @kib wrote:

The use of dedicated (sub)zone is reasonable for KVA-starved arches, where we do not want to waste 1MB of KVA per md. Instead KVA use was shared with other consumers.

It should still be 128KB on KVA-starved arches?

At the moment pbuf_zsecond_create() preallocates pbufs, so the change in fact reduces wasted KVA (until one has many vnode md devices).