Speed-up for qdecr_fastlm: Tweaks to code

[slamballais]

The slowest part of the vertex-wise regressions is the calculation of the standard errors:
s2 <- lapply(res, function(z) colSums(z^2 / df))
se <- lapply(1:m, function(z) do.call("cbind", lapply(s2[[z]], function(q) sqrt(diag(q * XTX[[z]])))))

It can be sped up:
se <- lapply(1:m, function(z) sqrt(tcrossprod(diag(XTX[[z]]), colSums(res^2)[[z]]) / df))

With 10,000 vertices of 1,000 participants, this is approximately the timing:
Unit: milliseconds
expr min lq mean median uq max neval
fun1() 98.1338 102.0863 106.89275 105.352 109.5212 152.5436 100
fun2() 24.2160 25.3097 30.44813 26.151 31.4312 150.1490 100

So from median 105.4 ms to 26.2 ms. The full loop went from 265.8 ms to 157.3 ms (~1.7x faster).

[slamballais]

This was implemented in #46