Aka remove linux blinders.

This reverts commit f81b172d.

Refs #2884

This makes it easier to start the UDS-listening peer, and/or set
permissions on the socket file, after starting Varnish or loading
VCL with a UDS backend.

Closes #2884

Fixes #2885

Apologies for the bad version in 1f32c159

Closes #2852

This reverts commit 1f32c159.

VUT_Arg('i') excludes all records with tags that are not explicitly
marked as included, so this change had the effect that only records
with Begin, End and Link, and any other tags that might be mentioned
for inclusion, get written to binary logs. For u00006.vtc in
particular, only those three record types were written to the log.

What we need here is that the three record types are never excluded,
in order to get transaction grouping right for reads from binary logs.

Closes #2852

Re-using struct vmod_priv helped simplicity, but had the following
shortcomings:

- in ddfda3d7 we documented what was
  long implied: return values of vmod functions are assumed immutable
  and so are blobs.

  So the blob pointer should be const, yet the struct vmod_priv pointer
  can't be.

- the struct vmod_priv free pointer implies that there would be
  automatic memory management provided by varnish core, yet this does not
  exist for a good reason: We would otherwise need to track all blobs
  ever returned by vmod functions/methods.

So we turn the data pointer into a const and remove the free function
callback.

We also add a type field, which is to be viewed similar to the miniobj
magic, except that it should not be asserted upon. The type field is
intended for additional (yet unreliable) checks of vmods using BLOBs
to carry around vmod-specific private data.

Not what I came here for, but...

Encoding and decoding are as much a function as handling blobs,
in particular with transcode().

Closes #2870

Addresses the doc part of #2872.

Solaris, (remember Solaris?) will return connection related errnos
on fd metadata operations such as ioctl(2), [gs]etsockopt(2) etc.

We propagated these errors up from VTC_(non)blocking() which could,
theoretically, bail out earlier than it would otherwise have done.
This complicated code far more than it speeded anything up.

Instead just ensure that any errors returned are indeed connection
related, and then ignore them, and let Solaris deal with the broken
connection same way as other operating systems.