A major bugfix for QPY's handling of loose circuit bits in c0ac5fb
(Qiskitgh-9095) caused the deprecated bit properties `index` and `register` to
be lost after the QPY roundtrip.  As far as Terra's data model is
concerned, the two circuits would still compare equal after this loss;
the output registers would still be equal to the inputs, and bit
equality is not considered directly since it general bit instances are
not expected to be equal between circuits.

Losing this information caused some downstream issues for the IBM
runtime, which was still relying on `Bit.index` working in some cases,
despite it issuing a deprecating warning since Terra 0.17 (April 2021).
While this can be corrected downstream, QPY can still do a better job of
roundtripping the deprecated information while it is still present.

The QPY format does not currently store enough information to
_completely_ roundtrip this information in cases that some but not all
owned bits from a register are present in the circuit.  (This partial
data is a decent part of the cause of the bugs that Qiskitgh-9095 fixed.)
Since this is just in the support of deprecated functionality that
Terra's data model does not even require for circuit equality (QPY's
goal), it seems not worth it to produce a new QPY binary format version
to store this, when the deprecated properties being removed would
obsolete the format again immediately.

This allows complete round-tripping of all the deprecated register+index
information in the `Bit` instances through QPY, restoring us to the
_intended_ behaviour before Qiskitgh-9095.  The behaviour in the dumper before
that did not allow full reconstruction, because some of the information
was lost for bits that were encountered in more than one register.

This fixes the dumper to always output all the indexing information for
all bits, not to use `-1` whenever a bit _is_ in the circuit but has
previously been encountered.  The `standalone` field on a register is
sufficient to tell whether the bits contained in it should have their
"owner" information set; it's not possible (in valid Qiskit code) to
have a register that owns only _some_ of its bits.  To accomodate this,
the register reader now needs to be two-pass.

…oundtrip

…oundtrip