The fastest non-MMX P54CS will run on any So7 board. Not So5 (without modification) as it needs the second multiplier pin for 3x multiplier. So that gives you a lower limit. As for the the upper limit, you say no AGP and no DIMM. That means you are limited to the latest chipset without SDRAM support. AT vs ATX is a motherboard thing, not a chipset thing, so just avoid ATX regardless.
Given So7 but no DIMM, you have a limited choice in chipsets:
ALi Aladdin III
Intel i430FX, i430HX
OPTi Viper-M
SiS 5511, 5571, 5586 (although the latter has integrated VGA, which you probably also don't want)
Via VP1
Out of that lot the i430HX is probably the all-round best, with support for:
- PLB cache
- 512MB EDO
- a second tagRAM for up to 512MB cached (otherwise 64MB as with other Intel chipsets)
- SMP (so dual So7)
- very high EDO performnce
The SiS 5571 comes a close second, with (very slightly) lower performance, no SMP and no 512kB caching.
Biggest non-chipset choice is async vs PLB cache. That makes a BIG (~15%) difference in performance, so if you want max performance, go PLB. If you prefer the 'authentic' old feel, go async - although by the time of the P200, PLB was starting to become the default, certainly for high-end systems like the ones a P200 would be used in.