More web documentation improvements

[rileyjmurray]

I'm opening this issue to brainstorm things we need for web docs.

Outstanding issues from the ICIAM web doc push

• The description of the core API for sparse sketching needs revision.
• We need to explain the idea of a major axis.
• Add a background page that explains row-major vs column-major layout.
• Add a "what is RandNLA" page
• Add a "quick-start" page that includes a minimal CMake project based on RandBLAS.

Other things that have occurred to me since the ICIAM talk

Web docs should have a "Why RandBLAS" page, that mentions things like the following:

• Efficient sparse sketching is critical in many applications. RandBLAS does a great job of it. Examples include ...
  • It can generate structured sparse operators. I don't know of light dependencies for "sample uniformly from an index set without replacement". So if one doesn't use RandBLAS, then you're probably going to implement the sampling algorithm yourself. Problem is: the standard implementation of Fisher-Yates (without workspace recycling) is cripplingly slow. So, not only does RandBLAS have this functionality, it's also well-optimized.
  • It's very flexible in how it lets you apply sparse operators. So even if you had a way to generate a sparse sketching operator for use in another sparse matrix library, the prospect of using that library might not be simple.
  • It's efficient in how it applies sparse operators. So even if you did have a method for generating the operator and even if you did need to apply that operator in some simple situation, you probably gain very little performance benefit over using RandBLAS' built-in function. RandBLAS is also portable in ways that, e.g., MKL is not.
  • Distributable dense sketching and only generating relevant submatrices is important. OpenMP parallelism is important. Reproducibility regardless of the number of threads is important.

[rileyjmurray]

Largely resolved by #83 (despite the name of that PR not indicating anything about web docs).