We use high quality tri-axial DC-accels ("instrumentation-grade") for in-vehicle (automotive industry) measurements of steering-response motions and vibrations. These accels cost approximately $1000-$1500 each. Some are MEMS based and come in a variety of packages approximately 1inch - 2inch in size. We have an application that requires hundreds of accels and are considering using inexpensive accels like Analog Devices model no ADXL335. Comparison of performance specifications between the ADXL335 and our instrumentation grade accels shows comparable performance, yet the price differential is large. We have also evaluated the frequency response of the ADXL335 vs the iinstrumentation grade on a calibration stand yielding favorable responses.  Why is the cost differential so large?, We understand that the ADXL335 is offered as a chip instead of the ruggedized, encased packaging of the instrumentation grade accels, but the cost differential still does not appear to be justified if the packaging and documented calibration are the only differences.  Is there a difference in the fundamental performance of "instrumentation-grade" accels and these ADXL335s?,  Are all MEMS-based accels similar in performance? We desire direct access to an expert who can discuss this comparison of the inexpensive MEMS accels(vs instrumentation grade). We have discussed this question with an application engineer at Analog Devices, who could not answer these questions and suggested that we post the questions at this site.
You did not specify which "instrumentation grade" accelerometers you compared to the ADXL335, but I'm not surprised that you see little difference in specified performance. The cost of individual calibration is rather high. The cost of packaging into ruggedized enclosures is also fairly high. But none of this seems to justify the hige price differential.
Simply put, the ADXL335 gets you 95% of the performance at 1% of the cost because we manufacture them in huge quantity almost totally by machine.
As for your direct questions:
1) Answered above.
2) Well "instrument grade" doesn't really tell me precisely whaich sensors you are refering to, but I'll say "not really". The very inexpensive plastic package used in the ADXL335 does limit performance a bit. Specifically long term bias stability. The plastic package is, like all plactics, hydrophillic (absorbs water). So as humidity changes the package can swell or shrink due to changing water content. This swelling and shrinking changes the mechanical stress on the MEMS element and, as a result, varys the null bias a little. All parts will be within data sheet specifications as long as the part is being operated in it's specified environmental range (guardbanding at test assures that), but you might see a few mg of bias drift day to day. Of course, you could always "zero" the bias output before a test run to get around this.
3) No. There are a few different ways of making a MEMS inertial sensors. Each has it's pros and cons. There are likewise many choices in the way you design the sensor (both mechanically and electronically), each of which will effect one parameter or another. For example, the recently announced ADXL362 uses very little power (under 1µA), less than any other accelerometer I know of. But it's bias stability over temperature is not as good as some other products Analog Devices offers. Here you have products built in a similar process than have different performance attributes.