Accelerate eth_evm_modexp by 25x by dividing input size by 8

[mratsim]

Found by @guidovranken in the Ethereum Foundation sponsored fuzzing campaign,

eth_evm_modexp is very slow on some input.

The issue is located here:

1. We read the length in bytes here:

   constantine/constantine/ethereum_evm_precompiles.nim

   Lines 425 to 428 in d0f4ad8

   	let
   	baseLen = cast[int](bL.limbs[0])
   	exponentLen = cast[int](eL.limbs[0])
   	modulusLen = cast[int](mL.limbs[0])

2. But we allocated as many 64-bit words as there were bytes instead of dividing by 8:

   constantine/constantine/ethereum_evm_precompiles.nim

   Lines 455 to 457 in d0f4ad8

   	template base(): untyped = baseBuf.toOpenArray(0, baseLen-1)
   	template modulus(): untyped = modulusBuf.toOpenArray(0, modulusLen-1)
   	template output(): untyped = outputBuf.toOpenArray(0, modulusLen-1)

The issue is that multiplication scales quadratically, about 1.5n² with the number of words. In our case the ratio before/after is 25x leading to a DOS vulnerability

[mratsim]

Exponentiation with even base and power-of-two modulus is now instant for a lot more moduli.

In particular the DOS vector has been accelerated by 2310x.

Before:

After generic fix (i.e. dividing input size by 8) - 28x acceleration.

Using special structure - 2370x acceleration.