The problem with direct control, such as switches (or relays) is that a fault in the driving circuit such as a stuck switch or relay will hold the coil on resulting in either a burning coil in EM games or fry the driving transistor in solid state games. CPU control avoids the problem, providing only a predetermined pulse time to the coil. Advancements allowed custom tailored variables to be added as operator adjustments (settings such as "coil pulse power" changes the pulse rate to the coil to adjust for varying line voltage conditions)

Having the bumpers be cpu controlled wasn't really a drain on the cpu resources. The switches need to be scanned either way, so just adding in a "oh, turn this coil on" to some of them isn't a lot more.

A lot of cpu issues came from ram/rom limitations more than processing speed limitations.

The early CPU controlled bumpers (bally/stern) use a faster scan for the switches that fire the coils.

Gottlieb and Williams stuck to a direct activated type setup, Gottlieb directly with a high power switch, and williams through logic chips.

In addition to other reasons mentioned above, I think the decision might have been influenced by manufacturing and maintenance costs as well. In the early Bally -17 design, they were going to have controlled coils for chimes/outhole/knocker, and they dedicated 4 bits to control 15 solenoid channels. The room was there for adding more controlled solenoids.

By making the pops and slings operate the same way, they only needed one set of contacts (Gottlieb uses one set of solenoid contacts and one set for scoring), and they also simplified the playfield to one type of contact (low-voltage only), which last longer and require less cleaning.

Solenoid voltage is also a little better isolated from logic voltage. A tech adjusting pop bumper relays is less likely to fry a control board.
Compared to Gottlieb, it really seemed like Bally was trying to minimize high-voltage relays and contacts. There's no Tilt relay, GI relay, and the only high voltage contacts are the flipper buttons / contacts.

Complete lost on most techs of the time, who happily filed the gold plating off at each shop job.

On early Data East pins the CPU had no direct control over the pops and slings and this was changed on the version 3 board. Phantom of the Opera was the last title to use the V2 board. What is curious is the drivers for the pops and slings are actually on the MPU in all versions so it is something they could have done from day one.

V1 and V2 pins can use the V3 board with no changes.

There is no other differences between the V2 and V3 boards and the code was literally at the whim of the individual programmer till the early 90's when they standardized a bit.

CPU speeds remained unchanged.

If I had to guess it was for safety reasons (cpu could see coil actuation) and control the power (adjustable on several titles)

Yeah, all that changed with DE was that they stopped hooking the switches directly to the transistors (danger of locking on), and instead scanned the pops via the switch matrix to fire them

Controlling bumpers and slings directly with transistors was an important feature of the early SS games. By using software to control the pulse to the bumper we were able to better control the strength and quality of the bumper's 'bump'.
By the mid 1970s Bally was already using DC rectifiers in EM games to make bumpers bump better by running them off of DC. Seemed like a no brainer to use the micro computer to control them. When Bow and Arrow prototype came out we had transistors controlling all coils including the flippers. We put SS games side by side in arcades with their EM counterparts to determine which was better. Not only did the SS game make more money players preferred the action of the bumpers to the EM games. We also found out having the flippers direct controlled with transistors was not effective and caused the drive transistors to burn out often. That's why production Bally games went with an enable relay.
The downside was coils locking up and frying transistors. The 1A playfield fuse was an attempt to contain that issue. The core problem at the time was CPU crash. The brown Palmer sockets Midway used to build the MPUs was the main issue.

I've seen this kind of action slow on older games, but the cause is usually that the switch contacts were dirty, or well worn and needed to be replaced.