per-layer feature mask

[nihui]

create a new param entry with id 31=uint
use bit for per-layer feature masking

bool use_fp16_packed;
bool use_fp16_storage;
bool use_fp16_arithmetic;

Sample use case

7767517
6 6
Input            data         0 1 data 0=224 1=224 2=3
Convolution      conv1_1      1 1 data conv1_1 0=64 1=3 4=1 5=1 6=1728 9=1
Convolution      conv1_2      1 1 conv1_1 conv1_2 0=64 1=3 4=1 5=1 6=36864 9=1
Pooling          pool1        1 1 conv1_2 pool1 1=2 2=2
Convolution      conv2_1      1 1 pool1 conv2_1 0=128 1=3 4=1 5=1 6=73728 9=1
Convolution      conv2_2      1 1 conv2_1 output 0=128 1=3 4=1 5=1 6=147456 9=1

Typically, we use fp16 computation to improve inference speed
Because the weight value of conv2_1 is large, fp16 accumulation may cause numerical overflow, so fp16 needs to be disabled individually for conv2_1, while other layers continue to use fp16 mode

Add 31=1 i.e. (1<<0) as disabled bit to disable fp16

Convolution      conv2_1      1 1 pool1 conv2_1 0=128 1=3 4=1 5=1 6=73728 9=1 31=1

It is also possible to control num_threads for each layer individually, but it is not very useful, so no more precious bits are used

mask	bit	rationale
no fp16 arithmetic	1<<0	precision concern
no fp16 storage	1<<1	precision concern
no bf16 storage	1<<2	precision concern
no int8	1<<3	debug dynamic quantized model
no vulkan	1<<4	reduce overhead for cpu op - gpu split - cpu op
no sgemm	1<<5	reduce some memory
no winograd	1<<6	reduce some memory

These masks will be implemented, and more bits can be used to achieve other needs in the future