I have begun learning about computer hardware many years ago on computers with 8-bit Intel 8080, Zilog Z80 and Motorola MC6809 CPUs and I was among those who could read hexadecimal Z80 machine instruction codes as easily as a source text in a high-level programming language, so I know from first-hand experience how it is to learn on 8-bit CPUs.
Nevertheless, given a choice, I would never use those again for this purpose.
None of the traditional 8-bit CPUs (i.e. introduced before 1980) had many instructions that took exactly one clock cycle. Most of their instructions required multiple clock cycles with a number different for each instruction. Nevertheless, on many 8-bit CPUs you could guess approximately the number of clock cycles for many instructions based on the number of memory cycles needed for executing the instruction (including the cycles required for fetching the instruction code).
On the contrary, modern 32-bit microcontrollers have a majority of instructions that are executed in one clock cycle. Regarding the cache memory, the slower models with a clock frequency below 100 MHz, e.g. from 24 to 72 MHz may have no cache, and even if they have a cache, not enabling it would not make much difference.
From all that you have said I agree only with the fact that a 32-bit MCU will have usually a complex clock tree with various PLLs that must be programmed for maximum performance.
Nevertheless, besides the fact that programming the PLLs is usually needed only in a single place, at startup, it is also optional for many applications. All 32-bit MCUs that I have seen start after reboot in a safe mode using an internal RC oscillator. The speed is low, but it works. You need to program the PLLs only if you want to reach the maximum speed by using an external quartz oscillator or resonator. The 32-bit MCUs may have a large number of complex internal peripherals but at least in the beginning it is possible to ignore their existence. Only the few peripherals that are used must be enabled and configured and the MCU vendors provide example programs that can be used before learning how to modify them.
8-bit disassembly does not have a lower instruction count but it always has a much higher instruction count for doing the same task, because each 8-bit instruction does less work.
The easiest to program in assembly language are the 64-bit CPUs, because the range of 64-bit numbers is big enough to use them in most applications without worries. Already with 32-bit numbers you must be careful to avoid overflows, while with 16-bit or 8-bit numbers you must be prepared to handle overflow at each operation. So on CPUs with narrower hardware instructions and registers, you must almost always write or use a library implementing operations with wider numbers, which makes the programs more complex than for 32-bit CPUs, because only seldom you can do an addition or multiplication etc. by just using one hardware instruction.
For the 32-bit ARM CPUs all the software tools that one may need are free. For most 8-bit CPUs the tools are either proprietary or of lower quality than for the 32-bit CPUs.
There are 32-bit ARM CPUs that are soldered together with support circuits on a very small PCB with DIP pins (e.g. various models of STM32 Nucleo-32 boards), which can be inserted in DIP sockets or in solderless breadboards, so DIP is not an advantage exclusive to obsolete 8-bit CPUs.
There are many such 32-bit microcontrollers that have high-current pins that can drive directly LEDs and most cheap development boards may include buffers to provide 5-V tolerant I/O pins on the headers mounted on the board.
While for the cheapest 32-bit development boards the user interface must be done on some PC to which the development board is connected via USB, or the board may be used through a console on a serial interface or through a remote shell over Ethernet (for the boards including an Ethernet RJ-45 connector), there are more expensive boards, e.g. around $50, to which it is possible to connect a display panel directly.
It is true that I have never used an Arduino, because every time when I looked at any model it appeared overpriced and without any advantage whatsoever over a Cortex-M board, so I could not understand why would someone waste time with it.
On the other hand, the costs of learning to program a development board with Cortex-M are close to zero and the experience is much more valuable.