AP_Declination: Updated compressed lookup value table. Located 9 more failure points due to truncation on the packed LUT.

AP_Declination: Found 7 of the 22 failure points and fixed them. The issue with those 7 was the way I originally packed the value array. The last delta values on some rows were getting truncated. The new number of failures is down to 15 - WIP.

AP_Declination: Updated test to run in 5 degree increments. Changed to only print failures and also print total pass vs total fail.

AP_Declination: The exception signs unpacking logic was incorrect. I was shifting a 1 value left y%8 which would have needed a the signs to be packed right to left. My packed signs byte was packed left to right (left most being the 0 position) so I needed to reverse the shifting.

AP_Declination: In order for the bitwise & operation to work, negative signs need to be represented by 1s instead of 0s. Changed so that negative signs are represented by a 1 in the packed signs array. Updated lookup logic to reflect this change.

AP_Declination: Added original array to test sketch. Added loop to validate new compressed array against original in 1 degree steps.

AP_Declination: Moved rows with large spikes in delta to their own unsigned exception rows. The signs are in a separate packed array.

AP_Declination: Implemented Delta Encoding and Run-Length Encoding. Added method to traverse the compressed array and return the lookup value based on the same lat_index/lon_index that was used before.

AP_Declination: Reduced the size of the lookup table to support -75 <-> 70 latitude range. This allows the datatype to remain int8_t.

when a user first connects with USB, and later switches to the
telemetry port without restarting we were getting zero for error_yaw
in the logs, as AHRS.get_error_yaw() was being called twice.

This ensures we give the last value after the counter is reset

explain the algorithm a bit more

AP_OpticalFlow - small bug fix to ensure init attempts to read the product id 3 times before giving up

this caused problems with random()

This adds a large amount of noise robustness to the compass offset
learning algorithm, at a cost of 120 bytes of memory. The changes are
based on a long discussion with Bill Premerlani.

we don't need to add count any more, as floating point maths doesn't
need to round up

this seems to work much better than the old algorithm, converging
faster and more accurately. Even better, it has no linkage to DCM, so
no possibility of nasty feedback effects

APM_RC - moved Force_Out0_Out1, Force_Out2_Out3 and Force_Out6_Out6 to APM_RC parent class because it's already implemented in the APM1 and APM2 child classes anyway

Filter - added simple LowPassFilter (simple but it's possible to make errors with simple stuff too so might as well have one)

AP_OpticalFlow - resolved compile error in example sketch (it could not find DCM.h but it's not required anyway)

this is equivalent to multiplying by m.transposed(), but is more
efficient

this means we only link this in once, rather than for every use of
matrix multiply, which saves us some flash space

We need to be careful not to put large pieces of code in template
headers, as if the operation is used a lot, it costs us a lot of code
space

this implements a much more accurate model of a compass using matrix
rotations, instead of trying to calculate components directly

AC_PID - added more paranoid checking that imax is positive in constructor, operator() and load_gains methods

this allows us to use a tested and optimised rotation method

this is the core method used to update the DCM matrix with a gyro
vector. Moving it to AP_Math allows us to have a test for it