Raised upper limit for Comba squaring

[czurnieden]

This PR replaces the comba-squaring cut-off MP_MAX_COMBA/2 with MP_MAX_COMBA.

Saw @tomstdenis taking part in the discussion over at #429 lately and remembered the curious results for the faster squaring algorithms (Karatsuba and Toom-Cook-3) where the cut-offs for both coincided with the upper limit of comba-squaring but not for comba-multiplying.

The upper limit for comba-multiplying is the size of min(a.used,b.used) but for comba-squaring it is a.used/2. Neither the actual code nor tests with inputs of the form -1 + 2^(n*MP_DIGIT_BIT) (all bits set in every limb with MP_MASK) for all MP_16BIT, MP_32BIT, and MP_64BIT showed a reason for that difference.

Also: running tune gives reasonable cut-offs now for the Karatsuba and Toom-Cook-3 squaring algorithms.

[czurnieden]

It is not a proper mathematical proof but I can make an exhaustive test with the fully filled limbs (intermediate results cannot get bigger than with that input).

`MP_DIGIT_BIT`	`max floor(log_2(2*_W + W1))`
15	30
28	63
31*	62
60	127

So, yes: the Comba cut-off for squaring can safely be set to `MP_MAX_COMBA`.

\* 31-bit cannot be defined by simply adding "-DMP_31BIT" to `LTM_CFLAGS` it needs manual editing of `tommath.h`. (I think that has been done intentionally but I cannot find the discussion about it quickly)

[czurnieden]

I'm trying to clean up my PRs (quite a mess), so:
Is it useful or can it go?