[patch] Benchmarking bits, constructor for Whole_dist

[Jules Bergmann]

This patch mostly contains benchmarks bits: new benchmarks for FFTM and 
vector-matrix multiply.  Additional cases for fast convolution. 
Parallelize some fast convolution cases.

However, the patch also contains a new constructor for Whole_dist that 
takes a number of subblocks.  Since whole distributions can't be 
subdivided, the number of subblocks must always be 1.  This is checked 
at runtime.

This allows maps to be constructed a bit more conveniently.  For 
example, to construct a map with the first dimension distributed by 
block and the second dimension not distributed, you can now write:

	typedef Map<Block_dist, Whole_dist> map_type;

	map_type map(num_proc, 1);

Previously, you would have to write:

	map_type map(num_proc, Whole_dist());

				-- Jules