Home page: training across two machines, a C++ program that trains, and what the split costs - #21

Adds E43 (plugrl-server#99) and E44 (plugrl-server#100) to the home page, in both languages. Merge after #99 and #100: the links point to their directories on main.

• Under the coverage grid: a paragraph and a figure. The quickstart pair also learns with its env clients on a second machine, a Windows laptop on campus Wi-Fi feeding a Linux workstation. Each side has six seeds, and they fall in one band. A run took 41-53 minutes, against 20 on one machine. The figure draws each side's band of seeds and no mean line, because the means differ by less than the spread within either side.
• 'The environment side is light': the non-Python sentence now cites E44. A C++ program with nothing but the standard library steps its own Pendulum and trains a policy through PlugRL, as well as the Python env client does. E2 only exchanged invented data.
• 'What the split costs': replaces E7's VM-to-host number with the two terms E43 measured between physical machines. The first is a fixed latency of about 3 ms here. The second is twice the observation's bytes divided by the link's bandwidth (18 MB/s here). A faster link shrinks the second term in proportion.

mkdocs build --strict passes, and the new section was checked in a browser.