Second part of #650

Solves half #650

The rust compiler DCE works in a different way so conditionals on `cfg!`
should not be used if the code in the branches may not compile.

While at it move the `use` statements where they are needed to avoid
warnings.

Fix #657

Document NASM dependency.

Otherwise cargo-bench would try to call the test runner with --bench and
criterion-only options, making the process fail.

- Don't do duplicate tx_type rdo with bottom-up partition rdo.
- Set tx_type  as DCT_DCT when skip==true during mode decision rdo.
- Add set_tx_size() if rdo_tx_size_type() is not called.

Unbreak encoding without producing the reconstructed video.

* Add code to reorder frames during encoding

* Use BTreeMap instead of VecDequeue to hold input frames

* Implement sign bias in MV prediction

* Disable reuse of learned prob and change ref frame semantics

Use LAST2 instead of ALTREF when the second reference frame is also
in the past

- Move the frame information in a separate struct
- Move the EncoderConfig in api.rs

Manually format parts that skip formatting but don't fix the excess length.

Implement ExactSizeIterator and FusedIterator.

70%-90% faster than before according to criterion.

And use it to compute SAD.

* Format SUBSIZE_LOOKUP slightly then skip it

* Use stable but more verbose way to skip rustfmt

* Fix inter for 12-bit input.

12-bit is a special case that rounds differently.

See 7.11.3.2. Rounding variable derivation process in the spec.

Performs a near-exhaustive-equivalent search of sse-optimum deblocking
level setup at speed <= 3.  Currently always uses SSE, psychovisual
(cdef-dist) comes next.

Above speed 3, levels are set by quantizer.

Rather than converting a level to a threshold decision value and testing values against the threshold,
convert values to levels and compare directly against the filter configuration level.
This will allow binning decisions by level in a radix sort.

The belief that a deblocking pass at level zero is a no-op is incorrect;
disable the deblocking of an edge when its filter level is zero

This reverts commit ef97e0ec.

* Call av1_idct4() from inside av1_idct8().

* Call av1_idct8() from inside av1_idct16().

* Use 2 multiply rotation by PI/4.

Adds the full_btr_sqrt2() function that computes a rotation by PI/4
 using 2 multiplies instead of 4.

* Last two frames are used as references and are labeled as LAST
and ALTREF.

* Add rules to include MV with different reference index in MV stack

* Make 2nd reference frame more distant from current frame

* Use correct reference frame in chroma motion compensation

* Enable multiple reference frames only at slower speed settings

* Add unequal quantizer assignment to temporally predicted frames

Allow Github to detect the license