Misc improvements in XMGRES solvers

* Offer iterated classical GS apart from modified GS for better parallel scalability. https://onlinelibrary.wiley.com/doi/abs/10.1002/%28SICI%291099-0887%28200001%2916%3A1%3C57%3A%3AAID-CNM320%3E3.0.CO%3B2-I
* Memory (Krylov basis, Hessenberg matrix, etc.) should only be allocated but not initialized in bulk. Otherwise, memory consumption sky rockets if, e.g., Krylov basis is of the same size of the linear system (i.e., GMRES without restart).
* `maxiter=` kw-arg semantics should be such that it prescribes an upper bound for all inner loop iterations. Currently, it does not mean that.