A frame transformation over axis=1

Every frame transformation firepanda exposes runs down each column. pandas lets `axis=1` run across each row instead, so `df.diff(axis=1)` is each column minus the one to its left and `df.ffill(axis=1)` carries a value rightwards along a row.

Today `axis=1` raises by name with the reason in the message, which is the right refusal and not the right end state. It affects `shift`, `diff`, `ffill`, `bfill`, `dropna` and `cumsum` and its three siblings, so it is one kernel shape and not one fix per method.

It is a real piece of work rather than a plumbing change. A column store transforms down a column by walking contiguous memory, and transforming across a row means touching one element in each of N columns, which is a gather with a different type at every step. The columns have to agree on a type before the operation means anything, which is the same widening question `DataFrame.agg_all` had to answer for the reductions and probably wants the same three rules.

Falls out of [#336](https://github.com/tamnd/firepanda/pull/336) and part of [#334](https://github.com/tamnd/firepanda/issues/334).

Start by reviewing the current axis=1 refusal and the related work in #336 and #334. Trace the shared transformation kernel and the existing DataFrame.agg_all widening rules, then verify that shift, diff, ffill, bfill, dropna, cumsum, and its three siblings support row-wise behavior with compatible column types.

Written by the indexing model from the issue text.