Previously it wasn't easy for users to get the Column information; this
has to be calculated from the Start field, but this isn't convenient.

Also the start column in "column x-y" was off by one, so fix that too as
a bonus.

ParseError.Error() did:

	msg := pe.Message
	if msg == "" {
		msg = pe.err.Error()
	}

That was fine, but not very useful for people wanting to access the
Message field themselves, especially since err isn't exported.

For example staticcheck does this to add in the filename, and the error
is lost, because Message is often blank:

	% staticcheck
	staticcheck.conf:5:0:  (last key parsed: "dot_import_whitelist") (compile)

We now only use the err field to determine which error usage to display
in ErrorWithUsage()

Previously it would be on the next line:

    % tomlv <(print 'a=')
    Error in '/proc/self/fd/11': toml: error: expected value but found '\n' instead

    At line 2, column 3:

          1 | a=
          2 |
                ^

Because the map keys were converted to []string,
rv.MapIndex(reflect.ValueOf(mapKey)) would panic as it's the "wrong"
type (string, instead of the custom string type).

Just keep track of the reflect.Value instead; don't really need to
convert it to strings.

Fixes #413

I don't know if this is just a temporary thing, but the existing docs link isn't working today :)

The result of multiple append calls can return two slices backed by the
same backing array. The bug only surfaces under specific conditions when
the capacity of parser.context is increased by more than 1 due to this
append call.

Would panic if the TOML had just one line without a newline at the end
of the file.

Keys were marked as "Undecoded" in the meta when using UnmarshalTOML
from the Unmarshaler interface. Unlike Primitive, I don't think this
makes sense as UnmarshalTOML() explicitly "decodes" the value.

Also mark all values below the table or array, and assume the parent
UnmarshalTOML() handled this correctly.

Fixes #425

Don't need the exception as my PR to improve it got merged.

I'm not entirely sure if I understand what's going on here, but this
seems to fix it.

Fixes #430

The map is in fieldsSub, so when iterating over fieldsDirect the length
check would think it's "last" and won't write the comma, whereas on the
second writeFields() call the test would fail and it would (wrongly)
write the comma.

So just sum the lengths.

Fixes #438