llama-bench: fix case where mmap and direct-io are turned on together

[taronaeo]

Ref: #20211 (comment)

As mentioned in the thread referenced, users may inadvertently enable both --mmap and --direct-io together and notice no performance difference.

This PR adds additional safeguards during flag parsing to ensure that both --mmap and --direct-io cannot have the same enabled state. Also, moved the ggml_backend_load_all call just after the flag parsing function for llama-bench to fail early if any of the flags has an error.

Test Cases

1. --mmap is enabled by default but the user specifies --direct-io 1

$ build/bin/llama-bench -hf ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M --direct-io 1 

...truncated...
common_download_file_single_online: using cached file (same etag): /Users/taronaeo/Library/Caches/llama.cpp/ibm-granite_granite-3.3-2b-instruct-GGUF_granite-3.3-2b-instruct-Q4_K_M.gguf
| model                          |       size |     params | backend    | threads | mmap | dio |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ---: | --: | --------------: | -------------------: |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    0 |   1 |           pp512 |       633.47 ± 44.96 |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    0 |   1 |           tg128 |         59.78 ± 0.94 |

build: a320bed33 (8323)

2. --mmap and --direct-io arguments do not have the same number of values

$ build/bin/llama-bench -hf ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M --mmap 1,0 --direct-io 1

...truncated...
common_download_file_single_online: using cached file (same etag): /Users/taronaeo/Library/Caches/llama.cpp/ibm-granite_granite-3.3-2b-instruct-GGUF_granite-3.3-2b-instruct-Q4_K_M.gguf
error: --mmap and --direct-io must have the same number of values

3. --mmap and --direct-io are turned on together

$ build/bin/llama-bench -hf ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M --mmap 1,0 --direct-io 1,0

...truncated...
common_download_file_single_online: using cached file (same etag): /Users/taronaeo/Library/Caches/llama.cpp/ibm-granite_granite-3.3-2b-instruct-GGUF_granite-3.3-2b-instruct-Q4_K_M.gguf
error: --direct-io cannot be enabled with --mmap

4. Valid benchmark run

$ build/bin/llama-bench -hf ibm-granite/granite-3.3-2b-instruct-GGUF:Q4_K_M --mmap 1,0 --direct-io 0,1

...truncated...
common_download_file_single_online: using cached file (same etag): /Users/taronaeo/Library/Caches/llama.cpp/ibm-granite_granite-3.3-2b-instruct-GGUF_granite-3.3-2b-instruct-Q4_K_M.gguf
| model                          |       size |     params | backend    | threads | mmap | dio |            test |                  t/s |
| ------------------------------ | ---------: | ---------: | ---------- | ------: | ---: | --: | --------------: | -------------------: |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    1 |   0 |           pp512 |       652.69 ± 12.18 |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    1 |   0 |           tg128 |         59.32 ± 0.24 |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    0 |   0 |           pp512 |        659.92 ± 5.79 |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    0 |   0 |           tg128 |         60.16 ± 0.13 |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    0 |   1 |           pp512 |        663.38 ± 0.52 |
| granite 3B Q4_K - Medium       |   1.44 GiB |     2.53 B | MTL,BLAS   |       8 |    0 |   1 |           tg128 |         60.15 ± 0.10 |

build: a320bed33 (8323)