Request: support >65536 pieces

[Bulat-Ziganshin]

65536 data+parity blocks limit is too low for some PAR3-style applications. I will be be glad to see it supported.

[catid]

Challenges:

(1) Error locator polynomial evaluation: Fast Walsh Transform on 2^32 words is not so fast? It would take 8.6 GB to represent 2^31 inputs...!

(2) Multiplication cannot be done via tables anymore, so it would get slower to calculate... Maybe Fp^2 would be better.

(3) The ErrorBitfield optimization class would be very important to implement but also very tricky.

(4) The current decoder wastes a lot of perf rounding up to the next power of two. I don't think that's going to fly when the data is this big!

(5) General memory handling issues...

[catid]

If applications require just slightly more than 2^16 blocks I could add e.g. 20 bit finite fields. I'd prefer to do that in response to a real demand rather than a precipitated need though.